feat(cuda): add ldmatrix integer dequantization

[jinzhen-lin]

Stacked on #137.

1. CUDA 13.4 adds an ldmatrix instruction that dequantizes INT4 to INT8 while loading data from shared memory. This PR adds full support for this operation in W4A8 GEMM. Thanks to @LilySu for pointing it out in Ptx94 ldmatrix s8 s4 dense w4a8 #85.

2. Analysis and hardware testing revealed broader support than the PTX documentation describes: all architectures supporting ldmatrix (SM75+) can execute the conversion through patched SASS, and the instruction also supports INT2-to-INT4 conversion. This PR reuses patch_cubin, compiling ldmatrix.sync.aligned.m8n8.{x2,x4}.trans.shared.b16 placeholders and then patching the generated SASS. This enables W4A8 and W2A4 conversion on supported hardware, including with CUDA versions before 13.4. Native PTX is used for INT4-to-INT8 on SM90+ with CUDA 13.4+.

3. Adds HUMMING_DISABLE_PATCH_CUBIN=1 to disable patched instructions and undocumented hardware formats, including Blackwell E3M4/E0M3 support. Integer dequantization falls back to software when native PTX is unavailable; documented native instructions remain enabled.

N = K = 8192, per-token activation × per-channel weight, default heuristics. Values are speedups over software dequantization, measured with alternating CUDA graph runs.

W4A8

M	SM75	SM80	SM86	SM89	SM90 H20	SM90 H200	SM120	SM121
1	1.01x	1.16x	1.00x	1.07x	1.28x	1.11x	1.08x	1.00x
16	1.01x	1.13x	1.00x	1.04x	1.06x	1.14x	1.02x	1.00x
64	1.00x	1.11x	1.07x	1.07x	1.02x	1.15x	1.06x	1.00x
256	1.05x	1.13x	1.01x	1.04x	1.00x	1.10x	1.02x	0.97x
1024	1.05x	1.18x	1.06x	1.14x	1.00x	1.06x	1.01x	1.01x
4096	1.07x	1.10x	1.02x	1.09x	1.00x	1.06x	1.01x	1.00x

W2A4

M	SM75	SM80	SM86	SM89
1	1.04x	1.16x	1.01x	1.06x
16	1.03x	1.15x	1.01x	1.04x
64	0.98x	1.09x	1.03x	1.07x
256	1.06x	1.07x	1.24x	1.05x
1024	1.08x	1.28x	1.14x	1.16x
4096	1.08x	1.12x	1.10x	1.07x