perf(turboquant): bump mlx-vlm to 0.6.0 for batched-TQ decode speedup

[popfido]

Summary

Bumps the mlx-vlm dependency from the 6f60ee4 commit pin to the 0.6.0 PyPI release. This is the upstream-released version of the TurboQuant decode fixes oMLX has been carrying, and it additionally lands the RHT value-kernel speedup (Blaizzy/mlx-vlm#1252) that directly accelerates the batched TurboQuant decode path wired up in #1547.

Why it's faster

In 0.5.0 the L=1 value-sum fast kernel was gated off whenever RHT was active (if not self.use_rht and ...) and fell back to an einsum on every decode step. RHT is on for any power-of-two head dim (e.g. Llama head_dim 64), so batched-TQ decode always took the slow path. 0.6.0 removes that gate and applies the RHT inverse on the kernel result (_value_rotate_inverse), so the Metal kernel runs under RHT. The value-sum cost scales with KV length, so the win shows at long context — exactly the regime where you'd reach for TurboQuant KV compression in the first place.

Validation

Re-ran the #1547 harness plus a long-context decode A/B on mlx-community/Llama-3.2-1B-Instruct-4bit (ctx=3584, two-point measurement so prefill cancels and only decode steps are timed):

scenario	0.5.0	0.6.0	Δ
batch TQ (B=4, masked decode)	165.1 tok/s	249.8 tok/s	+51%
single TQ (B=1, causal/fused path)	124.2	126.1	flat
batch fp16 (reference)	506.5	493.8	within noise

The gain is isolated to the masked batch-decode path; single-seq (fused causal kernel) and fp16 are unchanged — confirming the improvement is the value kernel, not measurement drift.

KV occupancy and accuracy are identical to 0.5.0 (TQ 0.312× fp16; 69% projected saving at 8K context; batch-vs-single token match unchanged).

Blast radius

• No oMLX code changes. TurboQuantKVCache.{decode,prefill}_attention and from_cache signatures are unchanged; the existing omlx/patches/turboquant_attention.py routing works as-is.
• transformers (already 5.8.1) satisfies 0.6.0's >=5.5.0; mlx-audio was already a transitive core dep, so no new core packages.

Test plan

• TurboQuant suite (-m turboquant): 32/32 pass on 0.6.0
• Full CI-equivalent suite (pytest -m "not slow and not integration"): 4877 passed, 37 skipped (only the pre-existing, CI-absent local dflash_mlx env failures remain — unrelated to this bump)
• feat(turboquant): batched KV-cache compression (single + batch), no worse than single #1547 memory/accuracy harness: occupancy + accuracy identical to 0.5.0
• Long-context decode A/B: +51% batch-TQ decode (table above)