[Don't Merge] KernelAgent-generated kernels - V2

[kaiming-cheng]

This PR introduced an updated version of kernel-agent-generated kernels on a difficult set of problem. Previously, KernelAgent failed to generate correct kernels on the 25 ops below. In this version, the average correctness score is 0.69, with 14 ops reached 100% correctness (56% pass rate). This PR serves as a reference for the generated kernel implementations.

Experiment Results

Average Correctness Ratio: 0.69
14 out of 25 kernels reached 100% correctness

Index	Operator	Correctness Ratio	Speedup vs Eager
1	_log_softmax_backward_data.default	0.0000%	N/A
2	_patched_sub_tensor	0.0000% (Failed to generate)
3	_softmax.default	0.0000% (Failed to generate)
4	add.Tensor	90.0000%	N/A
5	bmm.default	100.0000%	N/A
6	add_.Tensor	0.0000% (Failed to generate)
7	div.Tensor	100.0000%	N/A
8	fill_.Tensor	0.0000% (Failed to generate)
9	eq.Tensor	100.0000%	N/A
10	ge.Scalar	0.0000%	N/A
11	gt.Tensor	100.0000%	N/A
12	lt.Tensor	100.0000%	N/A
13	masked_fill.Scalar	100.0000%	N/A
14	max.dim	100.0000%	N/A
15	maximum.default	100.0000%	N/A
16	mean.dim	37.5000%	N/A
17	minimum.default	100.0000%	N/A
18	mm.default	100.0000%	N/A
19	mul.Tensor	57.8947%	N/A
20	pow.Scalar	100.0000%	2.0489x
21	reciprocal.default	100.0000%	N/A
22	std.correction	75.0000%	N/A
23	sum.default	100.0000%	N/A
24	sum.dim_IntList	56.2500%	N/A
25	where.self	100.0000%	1.1503x

[Jack-Khuu]

For tracking, can you share the commit hash these were tested on?

[Laurawly]

@kaiming-cheng We focus on testing fp16/bf16 dtypes for triton kernels on GPU. Pls reference this PR: #111