See python/cpython#23317

Raised in #930.

For an expression like `f"{one:{two:}{three}}"`, `three` is not in an f-string spec, and should be tokenized accordingly.

This PR fixes the `format_spec_count` bookkeeping in the tokenizer, so it properly decrements it when a closing `}` is encountered but only if the `}` closes a format_spec.

Reported in #930.

This is an obscure one.

`_ if 0else _` failed to parse with some very weird errors. It turns out that the tokenizer tries to parse `0else` as a single number, but when it encounters `l` it realizes it can't be a single number and it backtracks.

Unfortunately the backtracking logic was broken, and it failed to correctly backtrack one of the offsets used for whitespace parsing (the byte offset since the start of the line). This caused whitespace nodes to refer to incorrect parts of the input text, eventually resulting in the above behavior.

This PR fixes the bookkeeping when the tokenizer backtracks.

Reported in #930.

Python accepts code where `lambda` follows a `*`, so this PR relaxes validation rules for Lambdas.

Raised in #930.

This PR relaxes the accepted types for the `elt` field in `ListComp`, `SetComp`, and `GenExp`, as well as the `key` and `value` fields in `DictComp`.

Fixes #500.