Add LayerNorm Backward

[ighoshsubho]

This PR adds LayerNorm backward to QuACK. Instead of forking a new kernel, I extended the existing RMSNormBackward with an is_layernorm flag, since the two ops share the same memory pattern and only differ in the math at the row-reduction step. RMSNorm bwd needs one row reduction (mean(x_hat * wdy)) and writes only dW.

LayerNorm needs a second one (mean(wdy)), uses x_hat = (x - mean) * rstd instead of x * rstd, and additionally writes db = sum_rows(wdy). The dx formula becomes (wdy - mean(wdy) - x_hat * mean(x_hat * wdy)) * rstd, and db reuses the same (sm_count, N) fp32 workspace + host-reduce pattern that dw already uses, so we still avoid atomics. To carry two row reductions per row through the cluster pipeline, the reduction buffer goes from stage=2 to stage=4 (two slots × double-buffered for row pipelining). The rmsnorm path passes mMean=None and keeps the original 2-stage layout, so existing rmsnorm tests are untouched.

What's Added

• A public layernorm_bwd(x, weight, dout, rstd, mean) API that mirrors rmsnorm_bwd, returning (dx, dw, db). It wraps a torch.library.custom_op (quack::_layernorm_bwd) and requires fp32 weight (same constraint as the rest of the bwd path).
• Changed RMSNormBackward to accept is_layernorm, take an optional mMean tensor, and switch on the LN dx formula and second row reduction when set.
• Separated the reduction-buffer stage from the smem prefetch stage in the kernel so the LN path can run with stage=4 reductions and stage=2 prefetch without OOB writes.
• Added backward tests in tests/test_layernorm.py next to the forward tests, with the same (M, N, dtype, eps) parametrization (M ∈ {1, 37, 199}, N up to 262144, bf16/fp16/fp32). They validate against F.layer_norm autograd and skip the N > 128k && fp32 combo that RMSNormBackward already rejects on smem grounds.

Benchmark (B200)

M	N	dtype	quack ms	GB/s	torch ms	GB/s	speedup
64	256	bfloat16	0.0057	71	0.0792	1	13.87x
64	256	float16	0.0060	67	0.0807	1	13.49x
64	256	float32	0.0056	90	0.0779	3	13.99x
128	1024	bfloat16	0.0054	371	0.0789	10	14.59x
128	1024	float16	0.0059	341	0.0877	9	14.92x
128	1024	float32	0.0054	520	0.0790	20	14.72x
1024	4096	bfloat16	0.0101	2981	0.0843	299	8.37x
1024	4096	float16	0.0104	2880	0.0832	303	7.98x
1024	4096	float32	0.0125	4419	0.0867	581	6.94x
4096	8192	bfloat16	0.0533	3957	0.1058	1904	1.98x
4096	8192	float16	0.0544	3878	0.1059	1902	1.94x
4096	8192	float32	0.0783	5269	0.1552	2594	1.98x
4096	16384	bfloat16	0.1549	2725	0.2062	1953	1.33x
4096	16384	float16	0.1551	2722	0.2062	1954	1.33x
4096	16384	float32	0.1892	4360	0.3257	2473	1.72x
8192	16384	bfloat16	0.2992	2757	0.3927	2051	1.31x
8192	16384	float16	0.2988	2760	0.3917	2056	1.31x
8192	16384	float32	0.3648	4468	0.5986	2691	1.64x
32768	4096	bfloat16	0.2562	3163	0.4761	1692	1.86x
32768	4096	float16	0.2605	3111	0.4715	1708	1.81x
32768	4096	float32	0.3912	4130	0.6431	2505	1.64x
32768	8192	bfloat16	0.3741	4332	0.6942	2320	1.86x
32768	8192	float16	0.3820	4242	0.6917	2329	1.81x
32768	8192	float32	0.5600	5770	1.1174	2883	2.00x

cc: @tridao @GarlGuo @JackCharlesZhang would appreciate your eyes on this when you get a chance