[Issue]: [MIOpen] All FP64 RNN/LSTM Tests Failing with "RNN datatype must be float or half"

[chiranjeevipattigidi]

Blocking Latest Submodule Bump PRs:

• rocm-libraries: Bump rocm-libraries from 5454e9e to 7d2ed43 TheRock#4098 5454e9e ... 7d2ed43
• rocm-systems: Bump rocm-systems from 093b66c to c6da7be TheRock#4097 093b66c ... c6da7be

---

Summary:

• Multiple tests in Full/GPU_LSTM_FP64, Full/GPU_LSTM_dropout_FP64, and related suites are now failing.
• All errors show a similar pattern and log output.

Representative Error Logs:

MIOpen Error: rnn.cpp:333: RNNDescriptor: Bad parameter(s). RNN datatype must be float or half.
MIOpen Error: rnn.cpp:666: Data type mismatch.
MIOpen Error: rnn.cpp:614: Data type mismatch between descriptors
MIOpen Error: rnn.cpp:523: Data type mismatch between descriptors
unknown file: error: C++ exception with description "bad allocation" thrown in the test body.
Config requires 2223117779496 Bytes to write all necessary tensors to GPU. GPU has 26960986112 bytes of memory.
[  FAILED  ] Full/GPU_LSTM_FP64.DoubleTest/1 ...

Full Logs: https://github.com/ROCm/TheRock/actions/runs/23374902279/job/68134107578?pr=4098
https://github.com/ROCm/TheRock/actions/runs/23374816529/job/68134089834?pr=4097

[aalabsi-amd]

@chiranjeevipattigidi, @amd-hsivasun, @geomin12,
The reason for miopen failure is the change introduced in this PR: https://github.com/ROCm/TheRock/pull/4008/changes.
The failures are coming from tests that are not in the category that miopen should run "standard" instead from a bigger set not currently maintained "Full".
This PR should fix this: ROCm/TheRock#4120
Do we want to go with a revert of the 1st PR or a fix forward in the 2nd PR?

[Vsevolod1983]

#5750
PR which removes FP64 for LSTM

[aalabsi-amd]

#5766
addresses miopen failures

[aalabsi-amd]

This PR as well resolved the clang format issue which affected the jenkins ci: #5742

[chiranjeevipattigidi]

@aalabsi-amd
New regression in miopen tests: #5851