Refactoring to fix stacked graphs with NaNs.

Refactoring to fix stacked graphs with NaNs.

For stacked graphs, draw gaps for just the specific parts of series that are
missing data, and use interpolation to help ensure that the overall graph shape
doesn't get distorted for missing points.

This replaces #139 "Fix NaN handling for stacked graphs" which had
aggressively propagated NaNs to ensure that graphs don't get drawn with
incorrect stacking.

Fixes issue 446 - Improve handling of NaNs in stacked graphs.

Detailed changes:

* Remove layouts_.datasets, instead create point object arrays directly from
  gatherDatasets_(). This was necessary since the pre-stacked datasets
  aren't able to store the additional data needed for handling gaps.

* For stacked graphs, track yval_stacked/y_stacked separately from yval as point
  properties.

* Remove unstackPointAtIndex which is now no longer necessary, since the points
  keep the original yval. This helps ensure that the returned values keep their
  original NaN value and don't expose the interpolated stacked values used for
  drawing gaps.

* Remove evaluateWithError, evaluateLineCharts and Dygraph.seriesToPoints_ now
  handle error bars directly.

* In gatherDatasets_, remove the unconditional copy of rolledSeries, the
  refactoring appears to have fixed the bug related to zooming with
  error bars on. (I could reproduce it in the original code by zooming
  dygraphs/tests/custom-bars.html.) For date windows (horizontal zoom),
  use .splice() instead of manually coping.

Squashed commit of the following:

commit 56efaef6a50c737b99a8e4e6a36e55c83fde237f
Author: Klaus Weidner <klausw@google.com>
Date:   Mon Jun 3 16:50:16 2013 -0700

    Style fixes as per Robert's requests.

commit 83bb38a071f407560abcca53b7f6e07d6b0ca0e5
Author: Klaus Weidner <klausw@google.com>
Date:   Mon Jun 3 15:53:59 2013 -0700

    Add stackedGraphNaNFill option to control NaN handling.

    Add a test for this, and add a docstring to the stackPoints_ method.

commit e16e3c381d5cc7bec9896b31dddebc627a941c33
Author: Klaus Weidner <klausw@google.com>
Date:   Mon Jun 3 15:20:24 2013 -0700

    Revert name back to .addDataset, add docstring.

commit d3e66a60c84656364829e8a5bb83ed0e2b9fb403
Author: Klaus Weidner <klausw@google.com>
Date:   Mon Jun 3 15:15:19 2013 -0700

    Add more type annotations to make the flow of point data clearer.

    Unfortunately it's still not very clear, the logic is a bit convoluted.

commit 33db3d27501c8b5f2f84baa4dd92bf8d6f0a35a3
Author: Klaus Weidner <klausw@google.com>
Date:   Mon Jun 3 15:00:38 2013 -0700

    Rename fixPathAttrs_ to cleanPathAttrs_ as requested.

commit f39168785efcbacd1de9a6275aa02c8e34e4c5a7
Author: Klaus Weidner <klausw@google.com>
Date:   Mon Jun 3 14:58:26 2013 -0700

    Add comment explaining baseline, as requested by Robert.

commit 80348ed3c9e677fb5ff1ef393056e1e2bea1bbf7
Author: Klaus Weidner <klausw@google.com>
Date:   Mon Jun 3 13:16:51 2013 -0700

    Track stroke/fill in CanvasAssertions.

    For lineTo operations, set strokeStyle to undefined if the path is
    filled (not stroked), and set fillStyle to undefined if the path
    is stroked only.

    Fix assertions in missing_points tests, those were apparently expecting the
    inflated numbers.

    Remove the strokeStyle='#000000' hack in dygraph-canvas which is now
    no longer needed.

commit 4f47af82ac297d27dc0ac9d5c6fd256e00cccec5
Author: Klaus Weidner <klausw@google.com>
Date:   Mon Jun 3 11:57:56 2013 -0700

    fix CanvasAssertions which was calling .match wrong, remove now-redundant check.

commit 8c53e9e0f3d362e58204ff5a777c90ddb3b589d5
Author: Klaus Weidner <klausw@google.com>
Date:   Thu May 23 11:00:50 2013 -0700

    Restore hasOwnProperty.

commit 621cc37cfaf33821d04de2a94aea63021f7a9647
Author: Klaus Weidner <klausw@google.com>
Date:   Wed May 22 15:35:16 2013 -0700

    Fix all-NaN series stacking and Proxy return values.

    Ensure that all-NaN series get treated as zero for stacking
    purposes, and that nextPoint gets set to null past the last
    stackable point. We want interpolation, but not extrapolation.

    The Proxy class didn't return values from calls, breaking
    testCorrectColors since it didn't get pixel data as expected.

    Update the input data for testInterpolation to check corner cases,
    including an all-NaN series.

commit eb52ff1c1bbdd55312cb2881b5182731c91c155c
Author: Klaus Weidner <klausw@google.com>
Date:   Wed May 22 12:40:44 2013 -0700

    Refactoring to fix stacked graphs with NaNs.

    For stacked graphs, draw gaps for just the specific parts of series that are
    missing data, and use interpolation to help ensure that the overall graph shape
    doesn't get distorted for missing points.

    This replaces https://github.com/danvk/dygraphs/pull/139 "Fix NaN handling for
    stacked graphs" which had aggressively propagated NaNs to ensure that graphs
    don't get drawn with incorrect stacking.

    Fixes issue 446 - Improve handling of NaNs in stacked graphs.

    Detailed changes:

    - Remove layouts_.datasets, instead create point object arrays directly from
      gatherDatasets_(). This was necessary since the pre-stacked datasets aren't
      able to store the additional data needed for handling gaps.

    - For stacked graphs, track yval_stacked/y_stacked separately from yval as point
      properties.

    - Remove unstackPointAtIndex which is now no longer necessary, since the points
      keep the original yval. This helps ensure that the returned values keep their
      original NaN value and don't expose the interpolated stacked values used for
      drawing gaps.

    - Remove evaluateWithError, evaluateLineCharts and Dygraph.seriesToPoints_ now
      handle error bars directly.

    - In gatherDatasets_, remove the unconditional copy of rolledSeries, the
      refactoring appears to have fixed the bug related to zooming with
      error bars on. (I could reproduce it in the original code by zooming
      dygraphs/tests/custom-bars.html.) For date windows (horizontal zoom),
      use .splice() instead of manually coping.