[functorch] Use FuncTorchTLSBase instead of ThreadLocalDebugInfo (pytorch/functorch#314)
Fixes https://github.com/pytorch/functorch/issues/303
ThreadLocalDebugInfo stores a shared_ptr to some TLS that is shared
between ALL threads. The behavior of FuncTorchTLS should be that each
thread has its own TLS.
Test Plan:
- new test