Implement NLLLoss (contiguous, no reduction, forward only)

[hieule88]

• Added no reduction NLLLoss forward operation and kernel.
• Added driver test in the new format and gtest for forwarding NLLLoss .
• New API is guarded by MIOPEN_BETA_API macro.
• Compared to ROCm pytorch:

float16

op_name	dtype	size (N C D1 D2)	ignore_index	model	direction	ROCm pytorch	MIOpen HIP	Improvement
NLLLoss	float16	[1 2 2 2]	-100	random	fwd	21166	5866	3.61
NLLLoss	float16	[2 10 128 128]	3	random	fwd	39069	7004	5.58
NLLLoss	float16	[5 13 17 11]	5	random	fwd	20318	7679	2.65
NLLLoss	float16	[8 12 256 256]	-1	random	fwd	108630	19749	5.50
NLLLoss	float16	[8 16 512 512]	10	random	fwd	217693	76579	2.84
NLLLoss	float16	[16 21 512 512]	255	stdc	fwd	344451	192497	1.79

float32

op_name	dtype	size (N C D1 D2)	ignore_index	model	direction	ROCm pytorch	MIOpen HIP	Improvement
NLLLoss	float32	[1 2 2 2]	-100	random	fwd	23742	5510	4.31
NLLLoss	float32	[2 10 128 128]	3	random	fwd	39677	6862	5.78
NLLLoss	float32	[5 13 17 11]	5	random	fwd	19662	8017	2.45
NLLLoss	float32	[8 12 256 256]	-1	random	fwd	120486	27144	4.44
NLLLoss	float32	[8 16 512 512]	10	random	fwd	247707	113375	2.18
NLLLoss	float32	[16 21 512 512]	255	stdc	fwd	371121	275084	1.35

bfloat16

op_name	dtype	size (N C D1 D2)	ignore_index	model	direction	ROCm pytorch	MIOpen HIP	Improvement
NLLLoss	bfloat16	[1 2 2 2]	-100	random	fwd	24574	5600	4.39
NLLLoss	bfloat16	[2 10 128 128]	3	random	fwd	39980	6275	6.37
NLLLoss	bfloat16	[5 13 17 11]	5	random	fwd	20494	7822	2.62
NLLLoss	bfloat16	[8 12 256 256]	-1	random	fwd	110822	19962	5.55
NLLLoss	bfloat16	[8 16 512 512]	10	random	fwd	230461	74642	3.09
NLLLoss	bfloat16	[16 21 512 512]	255	stdc	fwd	371521	191911	1.94

• Average over all cases:

type	average
float16	3.66
float32	3.42
bfloat16	3.99

[kyeonghwanryu]

The performance looks good, enough to go to https://github.com/ROCm/MIOpen .
Read the codes of the other cases in NLLLoss like non-contiguous and backward. I'll give you a new ticket soon.
You can change your Jira ticket's status to done. Good work.