Support CL path for MaxUnpooling2d

[chunhuanMeng]

Previously, due to the lack of CL path, a lot of time was spent on format conversion, converting CF to CL, so adding CL path to optimize op performance.

[chunhuanMeng]

forward

dtype	shape	op before	kernel before	op with pr	kernel with pr
bf16	['bfloat16[4, 64, 2, 2]', 'int64[4, 64, 2, 2]', [128, 128]]	54.688us	43.716us	43.792us	32.212us
bf16	['bfloat16[4, 65, 1, 1]', 'int64[4, 65, 1, 1]', [128, 128]]	55.800us	44.456us	45.064us	33.080us
bf16	['bfloat16[8, 128, 1, 1]', 'int64[8, 128, 1, 1]', [128, 128]]	145.432us	126.584us	108.292us	88.904us
bf16_ChannelsLast	['bfloat16[4, 64, 2, 2]', 'int64[4, 64, 2, 2]', [128, 128]]	69.856us	42.656us	41.460us	35.832us
bf16_ChannelsLast	['bfloat16[4, 65, 1, 1]', 'int64[4, 65, 1, 1]', [128, 128]]	72.764us	34.700us	41.564us	35.868us
bf16_ChannelsLast	['bfloat16[8, 128, 1, 1]', 'int64[8, 128, 1, 1]', [128, 128]]	226.276us	118.804us	137.692us	119.124us
fp16	['float16[4, 64, 2, 2]', 'int64[4, 64, 2, 2]', [128, 128]]	55.468us	44.088us	43.288us	31.768us
fp16	['float16[4, 65, 1, 1]', 'int64[4, 65, 1, 1]', [128, 128]]	52.940us	41.944us	44.120us	32.260us
fp16	['float16[8, 128, 1, 1]', 'int64[8, 128, 1, 1]', [128, 128]]	146.588us	127.560us	108.132us	89.688us
fp16_ChannelsLast	['float16[4, 64, 2, 2]', 'int64[4, 64, 2, 2]', [128, 128]]	69.004us	33.612us	43.384us	37.904us
fp16_ChannelsLast	['float16[4, 65, 1, 1]', 'int64[4, 65, 1, 1]', [128, 128]]	73.116us	34.976us	43.904us	38.424us
fp16_ChannelsLast	['float16[8, 128, 1, 1]', 'int64[8, 128, 1, 1]', [128, 128]]	228.096us	118.688us	131.080us	118.856us
fp32	['float32[4, 64, 2, 2]', 'int64[4, 64, 2, 2]', [128, 128]]	59.688us	45.960us	45.364us	31.804us
fp32	['float32[4, 65, 1, 1]', 'int64[4, 65, 1, 1]', [128, 128]]	60.064us	46.200us	45.360us	31.884us
fp32	['float32[8, 128, 1, 1]', 'int64[8, 128, 1, 1]', [128, 128]]	162.416us	130.208us	126.508us	98.188us
fp32_ChannelsLast	['float32[4, 64, 2, 2]', 'int64[4, 64, 2, 2]', [128, 128]]	81.880us	34.816us	48.620us	34.620us
fp32_ChannelsLast	['float32[4, 65, 1, 1]', 'int64[4, 65, 1, 1]', [128, 128]]	89.652us	36.656us	41.604us	35.452us
fp32_ChannelsLast	['float32[8, 128, 1, 1]', 'int64[8, 128, 1, 1]', [128, 128]]	268.680us	130.152us	157.360us	120.520us

backward

dtype	shape	kernel	kernel
bf16_backward	['bfloat16[4, 64, 128, 128]', 'bfloat16[4, 64, 2, 2]', 'int64[4, 64, 2, 2]', [128, 128]]	24.404us	21.932us
bf16_backward	['bfloat16[4, 65, 128, 128]', 'bfloat16[4, 65, 1, 1]', 'int64[4, 65, 1, 1]', [128, 128]]	23.796us	21.712us
bf16_backward	['bfloat16[8, 128, 128, 128]', 'bfloat16[8, 128, 1, 1]', 'int64[8, 128, 1, 1]', [128, 128]]	51.832us	49.336us
bf16_backward_ChannelsLast	['bfloat16[4, 64, 128, 128]', 'bfloat16[4, 64, 2, 2]', 'int64[4, 64, 2, 2]', [128, 128]]	42.656us	44.336us
bf16_backward_ChannelsLast	['bfloat16[4, 65, 128, 128]', 'bfloat16[4, 65, 1, 1]', 'int64[4, 65, 1, 1]', [128, 128]]	42.780us	42.956us
bf16_backward_ChannelsLast	['bfloat16[8, 128, 128, 128]', 'bfloat16[8, 128, 1, 1]', 'int64[8, 128, 1, 1]', [128, 128]]	137.560us	137.064us
fp16	['float16[4, 64, 2, 2]', 'int64[4, 64, 2, 2]', [128, 128]]	23.580us	21.600us
fp16	['float16[4, 65, 1, 1]', 'int64[4, 65, 1, 1]', [128, 128]]	23.008us	21.764us
fp16	['float16[8, 128, 1, 1]', 'int64[8, 128, 1, 1]', [128, 128]]	52.324us	49.952us
fp16_ChannelsLast	['float16[4, 64, 2, 2]', 'int64[4, 64, 2, 2]', [128, 128]]	42.424us	42.724us
fp16_ChannelsLast	['float16[4, 65, 1, 1]', 'int64[4, 65, 1, 1]', [128, 128]]	42.852us	43.008us
fp16_ChannelsLast	['float16[8, 128, 1, 1]', 'int64[8, 128, 1, 1]', [128, 128]]	137.312us	137.548us
fp32_backward	['float32[4, 64, 128, 128]', 'float32[4, 64, 2, 2]', 'int64[4, 64, 2, 2]', [128, 128]]	21.704us	22.756us
fp32_backward	['float32[4, 65, 128, 128]', 'float32[4, 65, 1, 1]', 'int64[4, 65, 1, 1]', [128, 128]]	21.752us	22.828us
fp32_backward	['float32[8, 128, 128, 128]', 'float32[8, 128, 1, 1]', 'int64[8, 128, 1, 1]', [128, 128]]	61.292us	50.636us
fp32_backward_ChannelsLast	['float32[4, 64, 128, 128]', 'float32[4, 64, 2, 2]', 'int64[4, 64, 2, 2]', [128, 128]]	41.496us	41.684us
fp32_backward_ChannelsLast	['float32[4, 65, 128, 128]', 'float32[4, 65, 1, 1]', 'int64[4, 65, 1, 1]', [128, 128]]	42.996us	41.604us
fp32_backward_ChannelsLast	['float32[8, 128, 128, 128]', 'float32[8, 128, 1, 1]', 'int64[8, 128, 1, 1]', [128, 128]]	147.556us	138.736us