Improve the Verbosity class in python/verbosity_mgr.py

[oskooi]

Context

python/verbosity_mgr.py defines the singleton that fronts two C globals — meep::verbosity (src/mympi.cpp:147, default 1) and mpb_verbosity (libmpb src/util/verbosity.c, default 1) — exposed to Python by SWIG as mp.cvar.verbosity and _mpb.cvar.verbosity. It is the mechanism behind mp.verbosity(0) and the ~45 verbosity.meep > 0 / verbosity.mpb > 0 guards in simulation.py, solver.py, and visualization.py.

The class works for its dominant use, but reading it turns up five genuine defects plus several rough edges. None are caused by the API's shape, so the fix is targeted repair rather than a redesign: the public surface (verbosity(N), verbosity.meep, verbosity.mpb, get/set/get_all/reset, the comparison operators) stays as it is and no read site changes.

Defects

1. Subclass property leak. make_property() does setattr(Verbosity, name, property(...)) — the base class, not type(self). VerbosityForTest in python/tests/test_verbosity_mgr.py therefore grafts foo/bar permanently onto the production Verbosity class, and once reset() empties _cvars those properties raise KeyError.
2. Stale global level. Per-flag assignment (verbosity.meep = 2) never updates _master_verbosity, so get(), int(), repr(), and every comparison silently go stale.
3. Unhashable singleton. __eq__ is defined without __hash__, so Python sets __hash__ = None.
4. Phantom flags. __init__ re-runs on every Verbosity() call even though __new__ returns the existing instance, so a bare Verbosity() registers a throwaway _dummy under a manufactured name cvar_N that is then managed forever.
5. Late registration clobber an explicit setting. add_verbosity_var() only calls set() for the first flag, so mp.verbosity(0) followed by import meep.mpb leaves verbosity.mpb at the C default of 1 instead of 0.
6. copy.deepcopy(mp.Verbosity) corrupts the singleton. __new__ returns the live instance, then copy overwrites its __dict__ with copies of the cvars — after which mp.verbosity(0) writes to dead objects and the C globals never change again.

Rough edges

• The 0..3 range check is duplicated in set() and in the property setter; a non-integral level reaches the SWIG setter produces a confusing error.
• reset() calls _init() on the instance it is about to drop, mutating state that outstanding references still use.
• solver.py:28 re-registers mp.cvar as "meep" (a no-op) and then reads verbosity.mpb, a name that exists only because mpb.i happened to run first.
• Comparison dunders raise TypeError on non-numeric operands instead of returning NotImplemented.

Additions

VerbosityLevel IntEnum, temporary(level) context manager, __index__, __delattr__, __dir__, one shared _check_level() validator (TypeError for non-integral — which also makes np.int64 work, where it previously hit a confusing SWIG error), NotImplemented on non-numeric comparisons, and type hints.

Also: python/solver.py and python/mpb.i now use the already-public add_verbosity_var() instead of "constructing" a singleton that already exists; VerbosityLevel exported via meep.i, additional tests, @@ Verbosity.temporary @@ added to the doc template.