Move dtypes.canonicalize_value to C++.

[copybara-service]

Move dtypes.canonicalize_value to C++.

Since this function is primarily "apply table of handlers", we can speed it up both by moving to C++ both:

• the handler finding code, and
• some of the most common handlers, notably the identity function.

Notes:

• We retain the ability to register handlers in Python for novel types.
• We now want to call a weak key weak value cache from C++ when canonicalizing np.ndarray objects, so we add a C++ API for that also.
• Since calling canonicalize_dtype is now faster than skipping it with a bunch of isinstance checks, we add a number of additional types (Tracer, ShapeDtypeStruct) to the set of things that canonicalize_value knowns how to handle and call it unconditionally in places like Primitive.bind().