Add a Bottleneck engine to Patch.rolling

[d-chambers]

Bobine currently uses Patch.rolling(..., engine="numpy").mean() for centered smoothing of 2-D DAS patches. Bottleneck is already an optional DASCore dependency and is substantially faster for these moving reductions: on a representative 120 × 20,000 patch with a 20-sample distance window, the full operation was about 4.9× faster locally (0.057 s → 0.012 s).

DASCore exposes Bottleneck through dascore.utils.moving, but Patch.rolling currently accepts only the NumPy and pandas engines. Using moving_window directly is not a drop-in replacement because Patch.rolling has its own step, coordinate, history, full-window NaN, and even-window centering semantics.

Would it make sense to add engine="bottleneck" directly to Patch.rolling?

The compatibility contract I would expect is:

• mean, median, min, max, std, and sum preserve the existing NumPy/pandas output shape, coordinates, history, full-window NaN behavior, step/overlap handling, and centering placement, including even windows.
• apply either raises a clear unsupported-operation error for the Bottleneck engine or deliberately falls back to NumPy; an explicit error seems less surprising.
• Requesting engine="bottleneck" without Bottleneck installed raises a focused optional-dependency error rather than silently selecting another engine.
• Automatic engine selection could remain unchanged initially; explicit opt-in would make the first change small and predictable.

An implementation could add a _BottleneckPatchRoller alongside the existing NumPy and pandas rollers, reusing the common coordinate/history logic while dispatching built-in reductions to Bottleneck moving functions. The important part would be matching Patch.rolling semantics rather than inheriting the separate centering/boundary behavior of utils.moving.