EXL3: use branch-free trellis window extraction

[mrv777]

The hoisted EXL3 decoder branches on each window’s bit offset. Replace that branch with unsigned shifts and a masked carry, preserving zero-offset behavior and all floating-point operations.

A GPU regression checks 35,040 windows across every offset and rejects a broken zero-offset mask.

M1 Max validation:

• 19 GPU tests pass; all 82 original fixtures match output bits. Another 87 large cases pass full scalar-reference checks, and 33 cases preserve buffer guards and row padding. Scalar errors remain unchanged.
• Five batches of 20 alternating pairs give median GPU-time ratios of 1.08–1.47× across the original 12 mul1/3-bit cases. Longer repeats confirm the weakest case; controls stay near 1.00×. Memory pressure stays normal.
• Full build: 1,313 pass, two platform skips, only the existing GLM failure (GLM: fix sparse attention thread limit on M1 #704). Native build and golden checks have no unexpected differences.

Broader 1-bit inst3 prefill timings remain inconclusive. Other chips and end-to-end throughput are unmeasured.

#681 has merged; this PR contains only the branch-free follow-up. The follow-up commit contains this change. Raw logs remain local.

[ashhart]

Thank you, @mrv777. A shift and a masked carry in place of the two-way window is a neat change, and the reasoning holds at every offset: the carry sits below bit 16 from 1 to 16 and is masked at 0, and above 16 both forms keep the same bits of both words. Your commit landed in 1.0.9 under your name (2e8749c61), and the PR's head is merged into the history with no change to the tree (5d7d05a25), so GitHub shows this PR as merged.

What we ran, on an M3 Ultra and an M5 Ultra, against the build without your change:

• Tests. test-exl3 and test-exl3-gpu: 32 of 32 on both chips, your window check included (31 of 31 before it).
• Bits. The bench's parity check between the hoisted kernel and the baseline finds no differing bit on either chip.
• Speed, k = 6144, n = 25600, mul1, the hoisted kernel's mean GPU time: one row, 1.209 to 0.997 ms on the M3 Ultra (17.5% less) and 0.945 to 0.836 ms on the M5 Ultra (11.5% less); 16 rows, 3.755 to 3.404 ms and 3.035 to 2.608 ms. The unchanged baseline kernel moved by under 1%.