Blosc2

* Honor nested cparams properties in kwargs. E.g. you can do:


blosc2.save_tensor(a, "test.bl2", mode="w",
filters=[blosc2.Filter.TRUNC_PREC, blosc2.Filter.BITSHUFFLE],
filters_meta=[13, 0],
codec=blosc2.Codec.LZ4,
clevel=9)


without a need to build a proper `cparams` dict first.

* C-Blosc2 upgraded to 2.4.3. It should improve cratio for BloscLZ in combination with bitshuffle.

* Prefer pack_tensor/save_tensor in benchs and examples

* Remove the testing of packing PyTorch or TensorFlow objects during wheels build.

* New `pack_tensor`, `unpack_tensor`, `save_tensor` and `load_tensor` functions for serializing/deserializing PyTorch and TensorFlow tensor objects. They also understand NumPy arrays, so these are the new recommended ones for serialization.

* ``pack_array2`` do not modify the value of a possible `cparams` parameter anymore.

* The `pack_array2` / `save_array` have changed the serialization format to follow the new standard introduced in `pack_tensor`. In the future `pack_array2` / `save_array` will probably be deprecated, so please change to `pack_tensor` / `save_tensor` as soon as you can.

* The new 'standard' for serialization relies on using the `__pack_tensor__ ` attribute as a `vlmeta` (variable length) metalayer.

* Add `msgpack` as a runtime requirement

* New `pack_array2()` and `unpack_array2()` functions for packing NumPy arrays. Contrarily to `pack_array()` and `unpack_array()` counterparts, the new ones allow for compressing arrays larger than 2 GB in size.

* New `Scunk.to_cframe()` and `blosc2.from_cframe()` methods for serializing/deserialzing `SChunk` instances.

* New `Schunk.get_slice()`, `SChunk.__getitem__()` and `SChunk.__setitem__()` methods for getting/setting slices from/to `SChunk` instances.

* The `compcode` parameter has been renamed to `codec`. A `NameError` exception will be raised when using the old name. Please update your code when you see this exception.

* More doc restructurings. Hopefully, they are more pleasant to read now :-)