HIP: enable mfma mmq on gfx908 and gfx90a for select datatypes and shapes

[IMbackK]

This PR enables the MFMA path graciously provided by @deepsek on CDNA1/gfx908 and CDNA2/gfx90a devices where this is more performant than the current blas path.

This PR is fairly careful and only enables the path on datatype and batch sizes where tested to be more performant. It is likely that other data types would also benefit, and i will follow up with more at a later time.

Measurements

All mesurents on gfx908, pp1024

Performance change:

Model	Microbatch size	Test	t/s master	t/s mi100mfma	Speedup
llama 13B Q5_K_M	32	pp1024	141.71	385.04	2.72
llama 13B Q5_K_M	64	pp1024	214.31	575.49	2.69
llama 13B Q5_K_M	128	pp1024	388.50	707.50	1.82
llama 13B Q5_K_M	256	pp1024	653.29	777.11	1.19
llama 13B Q5_K_M	512	pp1024	850.12	847.13	1.00
llama 13B Q5_K_M	1024	pp1024	1137.55	1136.94	1.00
llama 13B Q8_0	32	pp1024	159.84	441.99	2.77
llama 13B Q8_0	64	pp1024	263.17	627.86	2.39
llama 13B Q8_0	128	pp1024	485.83	637.20	1.31
llama 13B Q8_0	256	pp1024	749.96	763.62	1.02
llama 13B Q8_0	512	pp1024	915.36	934.04	1.02
llama 13B Q8_0	1024	pp1024	1232.74	1216.02	0.99
llama 7B Q4_0	32	pp1024	465.36	1227.27	2.64
llama 7B Q4_0	64	pp1024	728.69	1615.49	2.22
llama 7B Q4_0	128	pp1024	1105.51	1850.48	1.67
llama 7B Q4_0	256	pp1024	1553.94	2158.73	1.39
llama 7B Q4_0	512	pp1024	1890.85	2220.39	1.17
llama 7B Q4_0	1024	pp1024	2115.34	2291.17	1.08
llama 7B Q5_0	32	pp1024	416.92	1016.73	2.44
llama 7B Q5_0	64	pp1024	757.01	1449.28	1.91
llama 7B Q5_0	128	pp1024	1125.59	1738.29	1.54
llama 7B Q5_0	256	pp1024	1567.06	2025.97	1.29
llama 7B Q5_0	512	pp1024	1875.81	2094.18	1.12
llama 7B Q5_0	1024	pp1024	1978.71	2153.81	1.09
llama 7B Q6_K	32	pp1024	312.87	775.24	2.48
llama 7B Q6_K	64	pp1024	688.96	1026.81	1.49
llama 7B Q6_K	128	pp1024	1064.16	1248.84	1.17
llama 7B Q6_K	256	pp1024	1484.92	1470.56	0.99
llama 7B Q6_K	512	pp1024	1731.99	1719.56	0.99
llama 7B Q6_K	1024	pp1024	1836.89	1835.06	1.00
llama 8B Q4_K_M	32	pp1024	301.40	1023.17	3.39
llama 8B Q4_K_M	64	pp1024	628.02	1426.31	2.27
llama 8B Q4_K_M	128	pp1024	1197.32	1814.15	1.52
llama 8B Q4_K_M	256	pp1024	1986.34	2217.39	1.12
llama 8B Q4_K_M	512	pp1024	2498.98	2490.22	1.00
llama 8B Q4_K_M	1024	pp1024	2726.83	2718.23	1.00

MFMA forced for all datatypes and batch sizes:

| model                          |     params | n_ubatch |                  t/s |
| ------------------------------ | ---------: | -------: | -------------------: |
| llama 7B Q4_0                  |     6.74 B |       32 |       1278.43 ± 3.66 |
| llama 7B Q4_0                  |     6.74 B |       64 |       1722.63 ± 4.00 |
| llama 7B Q4_0                  |     6.74 B |      128 |       2009.38 ± 5.06 |
| llama 7B Q4_0                  |     6.74 B |      256 |       2401.78 ± 2.37 |
| llama 7B Q4_0                  |     6.74 B |      512 |       2470.53 ± 7.83 |
| llama 7B Q4_0                  |     6.74 B |     1024 |       2571.22 ± 2.30 |
| llama 8B Q4_K - Medium         |     8.03 B |       32 |       1080.02 ± 2.33 |
| llama 8B Q4_K - Medium         |     8.03 B |       64 |       1553.58 ± 1.27 |
| llama 8B Q4_K - Medium         |     8.03 B |      128 |       2012.27 ± 0.61 |
| llama 8B Q4_K - Medium         |     8.03 B |      256 |       2286.34 ± 0.87 |
| llama 8B Q4_K - Medium         |     8.03 B |      512 |       2358.34 ± 1.46 |
| llama 8B Q4_K - Medium         |     8.03 B |     1024 |       2376.71 ± 0.72 |
| llama 7B Q5_0                  |     6.74 B |       32 |       1080.44 ± 1.48 |
| llama 7B Q5_0                  |     6.74 B |       64 |       1589.45 ± 0.79 |
| llama 7B Q5_0                  |     6.74 B |      128 |       1897.23 ± 4.16 |
| llama 7B Q5_0                  |     6.74 B |      256 |       2250.23 ± 1.01 |
| llama 7B Q5_0                  |     6.74 B |      512 |       2320.69 ± 1.50 |
| llama 7B Q5_0                  |     6.74 B |     1024 |       2342.06 ± 6.30 |
| llama 7B Q6_K                  |     6.74 B |       32 |        843.44 ± 2.82 |
| llama 7B Q6_K                  |     6.74 B |       64 |       1137.25 ± 2.91 |
| llama 7B Q6_K                  |     6.74 B |      128 |       1392.07 ± 0.57 |
| llama 7B Q6_K                  |     6.74 B |      256 |       1568.60 ± 0.38 |
| llama 7B Q6_K                  |     6.74 B |      512 |       1597.53 ± 0.40 |
| llama 7B Q6_K                  |     6.74 B |     1024 |       1620.83 ± 3.65 |
| llama 13B Q8_0                 |    23.57 B |       32 |        453.19 ± 0.44 |
| llama 13B Q8_0                 |    23.57 B |       64 |        674.65 ± 0.18 |
| llama 13B Q8_0                 |    23.57 B |      128 |        663.48 ± 3.70 |
| llama 13B Q8_0                 |    23.57 B |      256 |        739.34 ± 0.95 |
| llama 13B Q8_0                 |    23.57 B |      512 |        770.74 ± 1.21 |
| llama 13B Q8_0                 |    23.57 B |     1024 |        779.23 ± 1.25 |

[IMbackK]

side note: i think it might also be worth trying stream-k on GCN