[CUDA] Fallback QMM

[zcbenz]

This is an implementation of QMM that is supposed to work for all devices and options, it is slow and mainly serves a fallback for cases that we are not optimizing for.

On A100 it has only about half memory bandwidth of the pipelined version.

FP16xINT4 with K-major weights:

M	N	K	QMM (GiB/s)	CUBLAS (GiB/s)	QMM (TFlop/s)	CUBLAS (TFlop/s)	Speedup (x)
16	4096	4096	157.1	1388.4	8.69	22.04	0.39
16	16384	16384	288.1	1659.6	16.27	26.50	0.61
64	4096	4096	138.1	1295.7	28.28	80.41	0.35
64	16384	16384	227.4	1575.9	50.34	100.07	0.50

FP16xINT4 with N-major weights:

M	N	K	QMM (GiB/s)	CUBLAS (GiB/s)	QMM (TFlop/s)	CUBLAS (TFlop/s)	Speedup (x)
16	4096	4096	172.6	1590.6	9.55	25.25	0.38
16	16384	16384	346.6	1596.7	19.58	25.50	0.77
64	4096	4096	100.7	1389.3	20.63	86.22	0.24
64	16384	16384	187.9	1491.8	41.60	94.74	0.44

This PR does not implement 3/5/6-bit quants and arbitrary K size, which I'll do in followup PRs.

Also it takes a lot of RAM to build the kernels, and I have switched to large arm runner with swap enabled for running tests.