Thread safety infrastructure for shared IrContainer

Add std::shared_mutex to IrContainer to protect shared mutable state
during concurrent access from parallel compilation threads.

- IrContainer public methods self-lock (shared_lock for reads,
  unique_lock for writes)
- Fusion mutation methods (registerVal/Expr, removeVal/Expr,
  removeStatementsCreatedAfter) acquire unique_lock then delegate to
  lock-free ContainerMutator static methods, avoiding self-deadlock
  on nested calls (e.g., removeVal → removeExpr)
- ContainerMutator is a PIMPL struct defined only in fusion.cpp,
  keeping lock-free impl methods out of the header
- Remove kPhase2DisableParallelCompile guard, re-enabling parallel
  compilation now that the mutex is in place
- Delete dead IrContainer::copy() and IrContainer::swap() methods

Author: mdavis36