Avoid in-memory copies when filtering memmap signals in SequenceInterpolator

[Copilot]

Fancy-indexing a np.memmap with non-contiguous column indices materialises the entire selected dataset into RAM at init time, defeating the point of memory-mapping. This adds a signal_ids parameter to SequenceInterpolator (and PhaseShiftedSequenceInterpolator) with a tiered strategy to avoid that:

Filtering strategies

• Contiguous IDs (e.g. [3,4,5,6]) → strided slice view, zero-copy, still lazy
• Non-contiguous + uncached memmap → _data stays as the full np.memmap; sorted indices stored in _signal_ids; column filter applied inside interpolate() after the per-query row read — the full array is never materialised
• Cached or .npy → subset applied once at init (smaller allocation)

Usage

# Only signals [0, 7, 42] are ever loaded from disk per query
interp = SequenceInterpolator("path/to/data", signal_ids=[0, 7, 42])
assert isinstance(interp._data, np.memmap)  # still lazy
assert interp.n_signals == 3

Affected components

• _is_contiguous_range() helper added
• SequenceInterpolator.__init__: routing logic + _sorted_signal_ids attribute for subclasses to reuse without re-sorting
• SequenceInterpolator.normalize_init: subsets mean/std for the lazy path
• SequenceInterpolator.interpolate: applies _signal_ids after row read; fixes empty-array shape to use self.n_signals
• PhaseShiftedSequenceInterpolator: subsets _phase_shifts via _sorted_signal_ids; interpolate() uses original column indices when _signal_ids is set (required because np.take_along_axis demands matching shapes, which don't hold in the lazy case)