Reject embedding dims where dim*nbits isn't a multiple of 8 (instead of panicking)

[raphaelsty]

Problem

Found while testing #86: posting documents with embeddings whose dimension makes dim * nbits not a multiple of 8 (e.g. a 3-dim toy embedding with nbits=4) crashed the indexing worker with a panic:

thread '<unnamed>' panicked at next-plaid/src/codec.rs:393:
index out of bounds: the len is 1 but the index is 1

ResidualCodec::quantize_residuals computes packed_dim = dim * nbits / 8 (floor division) but the packing loop writes dim * nbits bits = ceil(dim*nbits/8) bytes. When dim * nbits isn't a multiple of 8, packed_dim is one byte too small and the loop indexes past the end of the row. The bit-packing format stores exactly 8 / nbits dimensions per byte (and decompress reads whole bytes), so the codec fundamentally only supports dims where dim * nbits % 8 == 0 — always true for real ColBERT dims (96, 128), but a panic for arbitrary inputs.

The panic happened deep inside a rayon worker, taking down the indexing task. Via the API (post-#104) it showed up as a panicked update batch rather than a normal error.

Fix

Validate the constraint at the top of quantize_residuals and return a descriptive Error::Codec instead of writing out of bounds:

unsupported embedding dimension 3 for nbits=4: dim * nbits (12) must be a multiple of 8

This is a defensive hardening — it does not change the on-disk format or behavior for supported dims. (Supporting arbitrary dims would require a packing/unpacking format rewrite for partial trailing bytes, which isn't needed for real models.)

Verification

• New unit tests: quantize_rejects_unaligned_dim_instead_of_panicking (dim=3 → Err, no panic) and quantize_accepts_byte_aligned_dim (dim=2, since 2*4=8 is valid even though 2 isn't a multiple of 8 — the guard checks dim * nbits, not dim).
• End-to-end through the API: re-running the original 3-dim/2-token /update now logs update.batch.failed … Codec error: unsupported embedding dimension 3 … with 0 panics, and the server stays healthy (and the failure surfaces in /health via Make index/centroids updates crash-safe and expose update progress in health #104's failed-status tracking).
• Full make ci-quick green (fmt, clippy -D warnings, tests).