Add ChainLogger, PyTorchLogger and LookupLogger

[shadeMe]

ChainLogger - Allows multiple loggers to be daisy-chained. The number of links in the chain is hard-coded as confection neither supports second-order resolution of registry functions nor does its parser correctly convert list primitives inside the same. This also interferes with pydantic validation, making the following syntax impossible:

[training.logger]
@loggers = "spacy.ChainLogger.v1"
chain = [ { "@loggers": "spacy.TestLogger.v1", "val": "random"}, { "@loggers": "spacy.ConsoleLogger.v3", "progress_bar": "train"}]

PyTorchLogger - Queries PyTorch stats and passes them to other loggers. We wanted to track CUDA memory statistics during training/distillation, but there was no straight-forward way to pass that information to the Weights and Biases logger, thus giving rise to logger chaining.

LookupLogger - Meant to be used to as a development tool when working with logger interop (wrote it for testing the above two additions).

I still need to add support for tracking arbitrary values in WandbLogger, but that'll come in a separate PR.

[svlandeg]

Nice!

The loggerX arguments aren't very satisfying as you noted, but I don't see an immediate way around this either.

One thing: we should add at least a few unit tests here. I see that we haven't been in the habit of doing so for spacy-loggers, but we really should. Checking console output etc is not straightforward, but we can at least check initialization from a config, both happy and non-happy path, etc.

[shadeMe]

Added a new utility method that will be needed for the upcoming PRs.

[svlandeg]

Tests are looking good, thanks!

Just had a few more minor nitpicks...

[svlandeg]

Should be good to merge once the tests go green :-)
Nice work!