Rework and enhance type hierarchy and generics

[thetorpedodog]

Context: single-cell-data/TileDB-SOMA#638 and single-cell-data/TileDB-SOMA#540

• Pulls basic collection impl into BaseCollection, allowing Collection to add the semantics of "no semantics". This mirrors the implementation in tiledbsoma.
• Makes Measurement and Experiment inherit from BaseCollection. Previous problems with this were due to missing __slots__.
• Adds generic parameters to Measurement and Experiment that allow implementations to specify the exact types they provide, saving a lot of casting down the road.
• Renames lots of TypeVars to be clearer about their purpose.
• Adds overloads to add_new_collection for better type inference.
• Tightens _Experimentish to only expect read accessors, not writers.

While these new changes add a bunch of generic slots to the base collection types and experiments and measurements, the experience from the perspective of a SOMA library user will be roughly the same. That is to say, it's a little scary here, but the end user will still see theimpl.Collection[ElementType]. Type inference when using composed objects is better as well:

some_exp = theimpl.Experiment.open(...)

obs = some_exp.obs
reveal_type(obs)
# BEFORE: somacore.DataFrame
#         (i.e., the type system doesn't know what implementation
#         of the abstract DataFrame this is; it only knows about
#         the bare minimum DataFrame properties)
# AFTER:  theimpl.DataFrame

ms = some_exp.ms
reveal_type(ms)
# BEFORE: somacore.Collection[somacore.Measurement]
# AFTER:  theimpl.Collection[theimpl.Measurement]

some_meas = ms["whatever"]
reveal_type(ms)
# BEFORE: somacore.Measurement
# AFTER:  theimpl.Measurement

some_meas.X
reveal_type(ms)
# BEFORE: somacore.Collection[somacore.NDArray]
# AFTER:  theimpl.Collection[theimpl.NDArray]

There is no change at runtime; the actual types of the objects remain the same, but autocompletion, type checking, and other tooling has a much better idea of what is going on.

---

To show what this looks like on the tiledbsoma side, the diff is pretty small, but the key part is in io.py, where the cast(tiledbsoma.Measurement, ms[whatever]) no longer needs to happen, since the type system already knows it’s a tiledbsoma.Measurement. While that is the only change there specifically, there will be corresponding improvements in user code.

And just to reiterate: runtime behavior is identical, and any code which works now will continue to work, but static type inference is significantly improved.