Columnwise quantize

[nastya236]

Column-wise quantization for tensors stored in a column-major layout. This is used when one or both inputs are passed transposed (in qqmm backward):

• nt layout: used in the VJP to compute dL/dx (the second argument is transposed)
• tt layout: used in the VJP to compute dL/dw (both argument are transposed)

Overview:
Input [M, K] M-major.

• Each thread processes group_size elements in a column
• Load to registers
• Compute scale -> store to shared memory (pad to avoid bank conflicts)
• Quantize -> store to shared memory (pad to avoid bank conflicts)
• Write scales to [M, K/group_size] K-major
• Write quantized values to [M, K/elements_per_byte] K-major

nvfp4 qqmm:

M	N	K	layout	diff %
16384	11008	4096	nn	6.3
16384	11008	4096	tn	15.9
32768	11008	4096	nn	0.9
32768	11008	4096	tn	7.8
16384	4096	11008	nn	6.3
16384	4096	11008	tn	20.1
32768	4096	11008	nn	14.5
32768	4096	11008	tn	11.4
16384	12288	4096	nn	2.4
16384	12288	4096	tn	23.8
32768	12288	4096	nn	1.3
32768	12288	4096	tn	15.3
16384	4096	12288	nn	6.4
16384	4096	12288	tn	20.0
32768	4096	12288	nn	8.9
32768	4096	12288	tn	24.9
16384	27648	5120	nn	11.7
16384	27648	5120	tn	15.9
32768	27648	5120	nn	0.7
32768	27648	5120	tn	9.4
16384	5120	27648	nn	10.6
16384	5120	27648	tn	14.2
32768	5120	27648	nn	2.3
32768	5120	27648	tn	21.7

Probably can be optimized further.
Note: fixed small bug in QQMatmul::output_shape + removed unused reorder

[awni]

This looks awesome! Let's merge it after the tests clear.