Add GPU attention tuned block-size defaults and dispatch

[dlwh]

Summary

• Add GPU attention tuned block-size lookup table and expose it for dispatch.
• Wire levanter.grug.attention to use inferred tuned block sizes by default and support partial user overrides.
• Add pallas_mosaic attention kernel module and package exports.
• Keep existing fallback behavior while enabling device-aware defaults for GPU tuning.

Testing

• Not run in this pass.

Notes

• Draft PR for iterative kernel/perf tuning follow-up work.

[claude]

Claude Code is working…

I'll analyze this and get back to you.

View job run