🐛 [Bug] perf gap reduce on timm/ViT

[zewenli98]

Bug Description

Compare the perf of Torch-TRT against ONNX-TRT:

In fp16:

1. Skipping constant folding of embedding layers doesn't affect engine size or latency or precision
2. Disabling linear decomposition + adding linear converter reduces ~15% latency
3. opt_level=3 or 5 get almost same latency
4. onnx-trt takes much longer in compile time
5. torch-trt is ~9% slower than onnx-trt

[zewenli98]

Layer  Diff (ms)    Diff %     Layer Name
--------------------------------------------------
5      -0.301008    -76.53    % _gemm_mha_v2_myl2_3
77     -0.299627    -75.32    % _gemm_mha_v2_myl2_75
37     -0.299243    -75.23    % _gemm_mha_v2_myl2_35
93     -0.298915    -75.08    % _gemm_mha_v2_myl2_91
85     -0.298421    -75.03    % _gemm_mha_v2_myl2_83
21     -0.298223    -74.97    % _gemm_mha_v2_myl2_19
13     -0.297811    -74.88    % _gemm_mha_v2_myl2_11
53     -0.297617    -74.86    % _gemm_mha_v2_myl2_51
69     -0.297692    -74.83    % _gemm_mha_v2_myl2_67
29     -0.297607    -74.82    % _gemm_mha_v2_myl2_27
61     -0.297447    -74.82    % _gemm_mha_v2_myl2_59
45     -0.296575    -74.52    % _gemm_mha_v2_myl2_43
98     -0.000804    -25.95    % __myl_ReshTranReshReshGath_myl2_96

Torch-TRT's gemm_mha_v2 kernel is much slower than ONNX-TRT's

[zewenli98]

With linear decomposition + adding linear converter in #3703, it reduces ~15% latency
With SDPA fix in #3700, it reduces ~9% latency.
Now it has almost the same latency as onn-trt.