opencl: tiled mul_mat with local memory for f16 and f32

[lhez]

This PR adds another variant of tiled matmul for f32 and f16. They also use local memory for tiling and pretty much follow the standard pattern. The main difference from #14535 is that tiles from both src0 and src1 are transposed.

On Adreno 830,

master

model	size	params	backend	ngl	test	t/s
qwen2 1.5B F16	2.88 GiB	1.54 B	OpenCL	99	pp512	145.45 ± 2.56
qwen2 1.5B F16	2.88 GiB	1.54 B	OpenCL	99	tg128	17.68 ± 0.14

this PR

model	size	params	backend	ngl	test	t/s
qwen2 1.5B F16	2.88 GiB	1.54 B	OpenCL	99	pp512	174.30 ± 10.32
qwen2 1.5B F16	2.88 GiB	1.54 B	OpenCL	99	tg128	17.73 ± 0.12