hipblaslt-bench performance improvements

[pghysels]

This improves performance of hipblaslt-bench when timing a large number of matmuls.

There are 2 main changes

• Initialization of A, B and C matrices is done on the device instead of the host. This also avoids the hipMemcpy from host to device. But when a host copy is needed for the correctness check, the device matrix is still copied back to the host.
• A memory pool to reuse memory allocated with hipMalloc/hipMallocManaged or hipHostMalloc. This might slightly increase memory use.

Both changes together significantly improve performance.
I have a test yaml input file with ~500K tests.
With these changes, the code runs in ~42 minutes, compared to 8 hours before (on MI300).

I initially used rocRAND, but then removed this dependency. I can add that back if it is preferred over my own naive random number generator.

This is based on a commit by @bethune-bryant

[jichangjichang]

Do we have the execution time of hipblaslt-test with/wo this PR?

[pghysels]

With these changes ./hipblaslt-test took

real    19m58.491s
user    35m32.051s
sys     0m38.180s

while the develop branch took

real    21m12.295s
user    114m49.692s
sys     2m34.752s

On an MI210, 16 OpenMP threads.

[bethune-bryant]

This is significantly speeding up my benchmark collection!
Thank you for getting this finished!