Add support for Metal.jl

[zhenwu0728]

Now our model can run on M-series Mac using GPU.

[zhenwu0728]

Benchmark:

macOS 14.5.0, Darwin 23.5.0

Toolchain:
- Julia: 1.10.4
- LLVM: 15.0.7

Julia packages: 
- Metal.jl: 1.2.0
- LLVMDowngrader_jll: 0.3.0+1

1 device:
- Apple M3 Max (13.840 GiB allocated)

Arch	N	min	median	mean	max	memory	allocs	samples
CPU	1024	683.167 μs	701.167 μs	739.829 μs	945.500 μs	308.56 KiB	3075	10
CPU	1048576	586.618 ms	593.765 ms	606.757 ms	714.189 ms	307.33 KiB	2996	9
CPU	33554432	19.917 s	19.917 s	19.917 s	19.917 s	307.33 KiB	2996	1
GPU	1024	260.176 ms	277.004 ms	281.977 ms	309.268 ms	5.04 MiB	171640	10
GPU	1048576	221.820 ms	256.475 ms	248.758 ms	274.850 ms	5.04 MiB	171920	10
GPU	33554432	599.450 ms	626.963 ms	673.170 ms	1.036 s	5.04 MiB	171914	8

Arch	N	Ns	min	median	mean	max	memory	allocs	samples
CPU	1024	32	1.834 ms	1.972 ms	1.997 ms	2.231 ms	1.80 MiB	3695	10
CPU	1024	64	5.022 ms	5.266 ms	5.223 ms	5.355 ms	5.65 MiB	4486	10
CPU	1024	128	16.857 ms	17.008 ms	17.039 ms	17.416 ms	20.50 MiB	5689	10
CPU	1048576	32	563.261 ms	563.884 ms	576.715 ms	678.273 ms	1.80 MiB	3616	9
CPU	1048576	64	629.369 ms	631.995 ms	659.017 ms	849.875 ms	5.65 MiB	4407	8
CPU	1048576	128	654.082 ms	654.977 ms	682.798 ms	876.487 ms	20.50 MiB	5689	8
CPU	33554432	32	18.848 s	18.848 s	18.848 s	18.848 s	1.80 MiB	3616	1
CPU	33554432	64	20.795 s	20.795 s	20.795 s	20.795 s	5.65 MiB	4407	1
CPU	33554432	128	21.136 s	21.136 s	21.136 s	21.136 s	20.50 MiB	5689	1
GPU	1024	32	234.488 ms	252.981 ms	264.093 ms	316.517 ms	5.38 MiB	180665	10
GPU	1024	64	265.827 ms	298.026 ms	295.401 ms	326.257 ms	5.91 MiB	199244	10
GPU	1024	128	309.454 ms	346.847 ms	344.838 ms	382.297 ms	7.19 MiB	241456	10
GPU	1048576	32	214.440 ms	248.960 ms	249.994 ms	268.858 ms	5.39 MiB	180913	10
GPU	1048576	64	263.491 ms	277.978 ms	285.221 ms	323.347 ms	5.92 MiB	199491	10
GPU	1048576	128	297.155 ms	331.759 ms	325.221 ms	351.346 ms	7.22 MiB	241840	10
GPU	33554432	32	546.911 ms	581.667 ms	613.843 ms	899.804 ms	5.39 MiB	181009	9
GPU	33554432	64	524.873 ms	589.794 ms	614.841 ms	871.087 ms	5.92 MiB	199572	9
GPU	33554432	128	597.942 ms	659.205 ms	694.217 ms	929.509 ms	7.20 MiB	241749	8

Arch	N	Ns	min	median	mean	max	memory	allocs	samples
CPU	1024	32	12.877 ms	13.145 ms	13.384 ms	15.230 ms	933.91 KiB	3984	10
CPU	1024	64	92.317 ms	94.492 ms	94.585 ms	96.147 ms	5.40 MiB	7697	10
CPU	1048576	32	657.237 ms	660.457 ms	688.958 ms	887.481 ms	933.91 KiB	3984	8
CPU	1048576	64	737.114 ms	748.572 ms	745.994 ms	756.733 ms	5.40 MiB	7697	7
CPU	33554432	32	21.287 s	21.287 s	21.287 s	21.287 s	933.91 KiB	3984	1
CPU	33554432	64	20.937 s	20.937 s	20.937 s	20.937 s	5.40 MiB	7697	1
GPU	1024	32	226.726 ms	257.110 ms	258.882 ms	327.908 ms	6.25 MiB	200722	10
GPU	1024	64	228.963 ms	261.200 ms	262.250 ms	297.629 ms	15.52 MiB	450049	10
GPU	1048576	32	184.881 ms	251.116 ms	252.394 ms	315.713 ms	6.24 MiB	200717	10
GPU	1048576	64	250.513 ms	300.799 ms	300.384 ms	343.170 ms	15.53 MiB	450547	10
GPU	33554432	32	546.412 ms	585.858 ms	635.806 ms	829.269 ms	6.25 MiB	200850	8
GPU	33554432	64	537.751 ms	560.650 ms	611.406 ms	830.820 ms	15.53 MiB	450770	9