Make `TensorBuffer::data()` non-virtual and move the pointer into the base class.

All existing `TensorBuffer` subclasses already store a pointer to
their buffer. Accessing that pointer by calling a virtual method is
inefficient. We currently generate the following instruction sequence
at the callsite (when compiling for x86_64 with a recent version of Clang):

1dd002:         mov    (%rdi),%rax                 tensor.h:655
1dd005:         callq  *0x10(%rax)                 tensor.h:655

...and the following implementation for `Buffer::data()`:

236520:         mov    0x10(%rdi),%rax             tensor.h:888
236524:         retq                               tensor.h:888

With this change, we generate a single `mov` instruction inline at the call
site, and avoid any branching.

PiperOrigin-RevId: 223985477