Fix dest→source multithread chunk sizing and add single-vs-multithread parity coverage

[Copilot]

Destination→source interpolation could fail in multithread mode when worker count exceeded destination points, due to mismatched partial block shapes during assembly. This update scopes block outputs to the active slice and adds explicit parity coverage to ensure multithread and single-core paths produce identical results.

• Root cause

  • In interpolate_block (dest→source mode), the block output buffer was allocated using full destination length instead of the current chunk length.
  • Empty/small trailing chunks then returned shapes incompatible with caller-side contiguous block writes.

• Code change

  • Allocate dest→source block output strictly by active slice length:

  block_size = max(chunk_idx.stop - chunk_idx.start, 0)
  interpolated = np.zeros([block_size, config.num_components])

• Regression coverage

  • Added a focused test that forces more threads than destination points and verifies dest→source interpolation no longer fails on empty trailing blocks.
  • Added a parity test asserting dest→source outputs (interpolated, unmapped) are exactly equal between multithread=False and multithread=True for the same input.

[dodu94]

seems ok, please add a test where it is demonstrated that multithread and single core interpolations yield exactly the same result