Fix qvm_split_k incorrect batch stride calculation - #3497

…plore#3497, 241f1bf)

…urces so they apply

The release scan added in magicnight#108 listed 33 upstream fixes the fork does not
carry. Triaging the three that touch quantized matmul — the path most users
actually run — found two that reach us and one that cannot.

ml-explore/mlx#3497: qvm mishandles a stride-0 batch dimension. That is exactly
the shape mlx-swift-lm's quantizedScaledDotProductAttention produces, where a
quantized KV cache is broadcast across GQA repeat groups with expandDimensions.
Only the non-transposed product (scores @ values) reaches the affected kernel,
and only for M >= 2, which is why ordinary single-token decode never showed it
while speculative decoding sits squarely in range. Not hardware-gated —
upstream reproduced on an M4 Pro. Measured here before the fix: max |diff|
0.139 at N=2048 and 0.104 at N=4096 against a dequantized reference, versus
~2e-4 quantization noise, with M=1 correct — matching upstream's numbers and
its stated boundary.

ml-explore/mlx#3631: NAX qmm edge-tile bounds computed in short(), wrapping
past 32767, corrupting a column band for unaligned N above 2^15. Reached
through an lm_head whose vocab is unaligned. MiniCPM3-4B's 73448 is one, and
one this project lists as supported; every other vocab we support is a multiple
of 64 and takes the aligned branch. Measured here before the fix: max |diff|
236 at N=73448, K=2560, M=16.

ml-explore/mlx#3854 is NOT carried: upstream states affine quantization is
unaffected, and every model this project loads is affine.

Bumping the submodule was not enough. mlx-swift does not compile the
submodule's Metal kernel headers — it embeds them, stringified, into
checked-in files under Source/Cmlx/mlx-generated, and the JIT compiles those.
After the bump both tests still failed with byte-identical numbers, because the
generated copies still held the pre-fix kernel text. The fork now regenerates
them. #3810 hid this trap earlier: it changed host C++, which is compiled from
source, so a plain submodule bump was sufficient there and looked like the rule.

QuantizedMatmulParityTests pins both, using upstream's own repro shapes, and
includes the M=1 case as a control so the M>=2 boundary is asserted rather than
assumed.