PR tensorflow#22073: Add a README file in the gpu_specs directory.

Imported from GitHub PR openxla/xla#22073

Copybara import of the project:

--
659afd527c8cbe9e67dca813cb182c13c41889ef by Dimitris Vardoulakis <dvardoulakis@nvidia.com>:

Add a README file in the gpu_specs directory.

--
9a8bf3a49eff2882c8fbbea60644ca1c8ad45333 by Dimitris Vardoulakis <dvardoulakis@nvidia.com>:

Improved the README.

Merging this change closes tensorflow#22073

PiperOrigin-RevId: 721916973

third_party/xla/xla/tools/hlo_opt/gpu_specs/README.md

@@ -0,0 +1,7 @@
+The specs in this folder are obtained by calling the appropriate override of
+`StreamExecutor::CreateDeviceDescription`, then turned into
+`GpuDeviceInfoProto`. (E.g., for NVIDIA GPUs, we call
+`CudaExecutor::CreateDeviceDescription`.) The specs are useful when compiling
+with the flag `--xla_gpu_target_config_filename`. Since a hardware generation
+may have several SKUs, a spec may not be identical to what we would get on a
+particular machine, but it will be "close enough".