Fix remove_artifacts sparsity handling for median/average modes (#3290)

[Leonard013]

Fixes #3290

Why

Using remove_artifacts(..., mode="median"|"average", artifacts=..., sparsity=...)
and then calling get_traces() raises:

ValueError: shapes (M,) and (N,) not aligned

In RemoveArtifactsRecordingSegment.get_traces (median/average branch), when a
sparsity mask is given the traces are sparsified to the masked channels:

trace_slice_values = traces[trace_slice]
if mask is not None:
    trace_slice_values = trace_slice_values[:, mask]

but the artifact template self.artifacts[label] stays full-channel. The
subsequent np.dot(trace_slice_values.flatten(), artifact_slice_values.flatten())
(and the later subtraction traces[trace_slice][:, mask] -= ... artifact ...)
therefore operate on mismatched channel counts. The sparsity block in __init__
only asserts that every label has a mask; it never restricts the template.

What

Restrict the template to the sparse channels at use time, right where the
traces are sparsified:

artifact = self.artifacts[label]
if mask is not None:
    # restrict the template to the sparse channels so that it aligns
    # with the sparsified traces (see issue #3290)
    artifact = artifact[:, mask]

and use this local artifact for the duration, the dot product, and the
subtraction.

Design notes:

• Both artifact-provisioning paths are covered. self.artifacts[label] is
  full-channel whether the template was passed in via artifacts= or computed by
  estimate_templates, so slicing at use time handles both uniformly.
• Serialization stays correct. The stored template and _kwargs["artifacts"]
  remain full-channel and sparsity remains full masks, so to_dict()/load()
  round-trips reproduce identical traces (verified). Storing a pre-sliced template
  in _kwargs instead would double-slice on reload — __init__ would try to index
  a full-length mask into an already-1-channel template (IndexError) — which is
  why the slice is applied at use time, not stored.
• No-sparsity path unchanged. When mask is None the template is used as-is.

Tests

Added test_remove_artifacts_sparsity covering mode="median" and
mode="average" with a sparsity mask and two labels (each masking a different
channel subset). It asserts that:

• get_traces() no longer raises,
• channels outside the mask are byte-identical to the clean traces
  (np.array_equal),
• channels inside the mask have the artifact subtracted (not np.allclose).

Before the fix this test fails at the np.dot with
ValueError: shapes (1200,) and (2400,) not aligned (600 samples × 2 sparse
channels vs 600 × 4 full channels); after the fix the whole file passes (2 passed).

I also verified (manually) the to_dict()→load() round-trip reproduces identical
traces, and the precomputed-artifacts path with sparsity — happy to add either as a
committed test if you'd like it locked in.

Note (out of scope)

The exact minimal snippet in the issue supplies a 30-sample artifact while the
default ms_before/ms_after imply a 105-sample window, so it additionally hits a
duration mismatch (105 vs 30) that is independent of sparsity. This PR fixes the
sparsity/channel-alignment bug (the subject of #3290); after the fix the channel
dimensions match (the error, if a mis-sized template is passed, reduces to the
pre-existing duration check).

---

Prepared with AI assistance (Claude Code): the reproduction, root-cause analysis,
fix, and test were AI-drafted; the reasoning and verification (including the rejected
store-pre-sliced approach that double-slices on reload) are documented above.