Add float16 head_dim 256 TK=1 SDPA vector kernel for M1/M2

[jmmarkiewicz]

On M1/M2, fp16 SDPA with head_dim 256 fails in the 2-pass vector kernel with Maximum threads per threadgroup is 896 but requested 960 when the 2-pass route is taken (around 1024+ keys on M1 Ultra), 32 × gqa × q_len falls between 928 and 1024, and the mask is a bool mask or none (causal compiles to its own _c pipeline and runs fine in my tests). It was reported in oMLX #4609 by @JerilynFranz, running MTP verify on an fp16 copy of Qwen3.8-27B; fp16 copies are common on M1 because bf16 is emulated there.

I reproduced it with synthetic tensors, and Qwen3.6-35B-A3B should hit it too (16/2 heads × MTP width 4 = 1024 threads), though I haven't run that model. #4643 added _tk1 fallbacks for bf16/256 and fp32/192 and 256 but not fp16/256, and its test only ran those shapes with mask="causal", so the gap never surfaced. This PR adds the missing instantiate_sdpa_vector_tk1(float16_t, 256) and the matching case in has_tk1_variant, and extends test_sdpa_vector_2pass_threadgroup_limit to run every shape with causal, no mask, and a random bool mask against the fp32 reference. _tk1 is only selected when the default kernel would exceed the limit, which used to throw, so paths that already worked are unchanged.

Closes #4718.

Testing (Mac Studio M1 Ultra, applegpu_g13d, macOS 27.2)

• Stock 0.32.4 repro: 4/36 failures, all fp16/256 with bool or no mask; fp16/192 never fails. At 256 keys the same shapes run fine.

• test_sdpa_vector_2pass_threadgroup_limit on main: before, 4 errors (exactly the fp16/256 no-mask and bool subtests); after, OK on all 24 subtests (8 shapes × 3 masks).

• Full test_fast_sdpa.py (32 tests): main 8 errors, this branch 4 errors. The remaining 4 are the existing test_sdpa_vector_head_dim_512 failures on M1: the 1-pass d=512 kernel is capped at 640 threads (fp32) and 832 (fp16), but MLX requests 1024. They're identical with or without this PR.

• ☑️ I understand it is strictly prohibited to use AI to write PR description

• AI usage disclosure: The root cause was found, and the patch and test written and validated, with Claude Code; I reviewed all changes.