Jzhou/3dconv

[jfactory07]

Proposed changes

Enhance 3D convolution performance in immediate mode (required for dynamic shapes to avoid kernel search overhead).

1. Smarter kernel selection: Optimizes fp16/bf16 performance by choosing better CK kernels contextually, rather than using the first kernel by default.

2. Runtime control: Adds the CK_CONV3D_IDX env variable for explicit kernel selection.

Checklist

Please put an x into the boxes that apply. You can also fill these out after creating the PR. If you're not sure, please don't hesitate to ask.

• I have added automated tests relevant to the introduced functionality
• I have sufficient test coverage for the changes, and code coverage hasn't decreased as a result of my PR
• I have ran the tests, and they are all passing locally
• I have added relevant documentation for the changes
• I have removed the stale documentation which is no longer relevant after this pull request
• I have ran make format & make check_format to ensure the changes have been formatted

[amd-hsivasun]

Imported to ROCm/rocm-libraries