compiler: hip: change default to use old driver for rocm 7.13

[ronlieb]

Restores more efficient processing of rocgdb call stack test.

flags to choose , default with this PR goes back to --no-offload-new-driver
--offload-new-driver
--no-offload-new-driver

ROCm 7.14 we will attempt to reenable new driver.

[searlmc1]

LGTM

[skganesan008]

rerun of rccl test failure shows the following:
Expected NULL handle for buffer size 4194304 bytes
[ INFO ] Test 'VariableSizeBuffers_Disabled' (PID: 5272) FAILED with exit code 1 after 2854 ms
[ INFO ] Running isolated test 'VariableSizeBuffers_Enabled' (PID: 5284) with env: NCCL_LOCAL_REGISTER=1
[ INFO ] Test 'VariableSizeBuffers_Enabled' PASSED (2857 ms)
[ INFO ] Running isolated test 'DeregisterNullHandle' (PID: 5296)
[16:40:13Z] Mem: 97.1/3023.4GB (3%) | CPU: 0% | Jobs: ~1/384 | Disk: 1341GB free
[ INFO ] Test 'DeregisterNullHandle' PASSED (2951 ms)
[ INFO ] Process-Isolated Tests: 4 passed, 3 failed, 0 skipped (20186 ms total)
[ INFO ] Failed: CommRegisterDeregister_Disabled - Test failed with exit code 1
[ INFO ] Failed: MultipleBufferRegistration_Disabled - Test failed with exit code 1
[ INFO ] Failed: VariableSizeBuffers_Disabled - Test failed with exit code 1
/_w/TheRock/TheRock/rocm-systems/projects/rccl/test/RegisterTests.cpp:244: Failure
Value of: passed
Actual: false
Expected: true
One or more isolated tests failed
[ FAILED ] Register.ProcessIsolatedRegisterTests (20189 ms)
[----------] 1 test from Register (20189 ms total)
[----------] 6 tests from Scatter

The same error is also seen on another open PR #4344

All of the pytorch build failures seems to be related to following error:
2026-04-04T05:33:16.6486344Z /opt/rh/gcc-toolset-13/root/usr/libexec/gcc/x86_64-redhat-linux/13/ld: /__w/TheRock/TheRock/external-builds/pytorch/pytorch/build/lib/libtorch_hip.so: undefined reference to `rsmi_init'
2026-04-04T05:33:16.6486962Z collect2: error: ld returned 1 exit status

This error is also seen on already merged PR #4337

[skganesan008]

We are landing the PR as the errors seen is not related to this PR based on the findings so far