[xla:ffi] Add benchmarks for internal XLA FFI implementation

[copybara-service[bot]]

[xla:ffi] Add benchmarks for internal XLA FFI implementation

• optimize forwarding absl::OkStatus() as nullptr XLA_FFI_Error* (success)
• make XLA_FFI_ATTRIBUTE_ALWAYS_INLINE consistent with ABSL

(1) xla/ffi:ffi_test

---

Benchmark Time CPU Iterations

BM_AnyBufferArgX1 19.5 ns 19.5 ns 35774168 BM_AnyBufferArgX4 52.5 ns 52.5 ns 13172198 BM_BufferArgX1 15.8 ns 15.8 ns 44010886 BM_BufferArgX4 34.1 ns 34.1 ns 20746753 BM_BufferArgX8 64.5 ns 64.5 ns 10999725 BM_TupleOfI32Attrs 65.8 ns 65.8 ns 10447148

(2) xla/ffi/api:ffi_test

---

Benchmark Time CPU Iterations

BM_AnyBufferArgX1 12.0 ns 12.0 ns 58216698 BM_AnyBufferArgX4 18.8 ns 18.8 ns 38334028 BM_BufferArgX1 12.3 ns 12.3 ns 49226385 BM_BufferArgX4 19.1 ns 19.1 ns 37307319 BM_BufferArgX8 37.4 ns 37.4 ns 18814343 BM_TupleOfI32Attrs 67.2 ns 67.2 ns 10336533