Bump hyparquet-writer to 0.16.6 for the dictionary-encoding fix

[philcunliffe]

Bumps the optional hyparquet-writer dep 0.16.5 -> 0.16.6, the release with the dictionary-encoding fix (hyparam/hyparquet-writer#35): the writer now keeps any dictionary that at least halves the encoded bytes instead of busting to PLAIN at the 1 MiB page cap. Server-side counterpart already merged (hyparam/hypaware-server#309); this covers the client's cache-tier writes.

The override: icebird 0.8.17 pins hyparquet-writer at exactly 0.16.5, and the cache write path goes through icebird, so a direct bump alone would leave a nested 0.16.5 where it matters. The override is scoped to icebird only; hypvector keeps its own 0.16.1 pin untouched. Drop the entry when icebird releases with its own bump.

Test updates: two format-parquet-clustering tests asserted the old writer's dictionary bust (it was the documented motivation for row-group clustering). They now assert the fixed behavior: the wide per-conversation column stays RLE_DICTIONARY with or without clustering, and clustering costs no material size (its remaining jobs are the peak-heap byte cap and read locality, per the updated rationale comment in the plugin). The old bust shape is kept as a regression canary.

Validation: typecheck green; 3,939/3,942 tests pass. The one failure (leave tears down a central layer whose active-slot pointer does not resolve (#623)) fails identically on unmodified master in two separate local checkouts, so it is pre-existing and unrelated (possibly macOS-specific).

🤖 Generated with Claude Code