Mixed-precision training: improve checks

[danieldk]

1. Before this change, using gradient scaling on a non-CUDA tensor would trigger an assertion. However, it is possible to trigger this error outside Thinc by enabling mixed-precision training on a CPU. So this should be a proper exception rather than an AssertionError. Besides raising a ValueError, the error message is also extended to describe how the error can be avoided.

2. We checked that gradient scaling is supported when construction a PyTorchGradScaler. However, this gives issues when someone uses a model that was trained with gradient scaling. In such cases, it's safe to contruct a grad scaler, since it is not used. This change moves the check to the actual scaling.

3. Also remove the check that verifies that mixed-precision scaling is available (when enabled) from the constructor of PyTorchShim. PyTorch will at most give a warning when trying to autocast when there is no support.

[danieldk]

Tested PyTorch versions:

CPU: 1.6.0 1.7.0 1.8.0 1.8.1 1.9.0 1.10.0 1.11.0
CUDA: 1.7.0+cu110 1.8.0+cu111 1.8.1+cu111 1.9.0+cu111 1.10.0+cu113 1.11.0+cu113

(All also with python -c "import spacy; spacy.load(\"en_udv25_englishewt_trf\")" to verify that we can load a mixed-precision model on the GPU.)

[danieldk]

Related issues: explosion/spaCy#10527 and explosion/spaCy#10543