Adopt the torch_tpu tensor_buffer API and drop the nobuilder build exclusions.

[copybara-service]

Adopt the torch_tpu tensor_buffer API and drop the nobuilder build exclusions.

The torch integration compiled against three torch_tpu internal headers
that only a source checkout provides, so every torch-dependent target
carried a "nobuilder" tag. tensor_buffer now covers everything the
integration uses.

• Tensors are resolved, inspected, and materialized through
  GetBaseTensorBuffer / TensorBufferHandle; every keep-alive holder
  stores the handle instead of DeviceBufferRef.
• BUILD: targets depend on tensor_buffer through a header_only_cc_info
  wrapper; the three internal-header forwarding targets and the
  "nobuilder" tags are removed. Exception: kv_cache_manager_test_py still
  imports torch_tpu._internal Python modules with no public equivalent,
  so it keeps "nobuilder" and is tagged "manual".
• The external wheel build (build.sh) passes --define
  raiden_wheel_build=true, which makes that wrapper header-only: the
  torch extension is compiled against the API but does not link its
  objects, and binds the symbols at import to the installed torch_tpu
  wheel, which exports them from its per-torch-version common library.
  Every other build links the API as a normal dependency. Previously the
  extension embedded its own copy of the API together with its allocator
  state, and a running torch_tpu rejected the extension's tensors on the
  allocator-identity check.
• The Copybara transform that stripped the device_buffer.h include from
  torch_tpu_utils.cc on export is removed; the include no longer exists.

Validation:

• Both torch packages build with no torch_tpu source checkout
  (26 + 25 targets), and the package unit tests pass on CPU (4/4).
• Wheel: the torch extension leaves the seven torch_tpu API symbols
  undefined in all three torch ABI variants (nm); built without the
  wrapper, they are embedded.
• On a TPU host, against a torch_tpu wheel that exports the API, the
  extension resolves GetBaseTensorBuffer from torch_tpu's common library
  and constructs a KVCacheManager over device tensors. The same probe
  with the previous link model aborts on the allocator-identity check.