Inspired by bpo-41064, I sat down to try and find problems with f-string locations in the new parser. I was able to come up with a way to compute the locations of the f-string expressions that I think is more consistent and allows us to delete all the code that was fixing the expression locations after the actual parsing, which accounted for about 1/6 of string_parser.c.

A high-level explanation of the change:

Before this change we were pre-feeding the parser with the location of the f-string itself. The parser was then parsing the expression and was computing the locations of all the nodes based on the offset of the f-string. After the parsing was done, we were identifying the offset and the lineno of the expression *within* the fstring and were fixing the node locations accordingly. For example, for an f-string like `a = 0; f'irrelevant {a}'` we were doing the following:

- Pre-feed the parser with lineno=0 and col_offset=7 (the offset of the f-string itself in the current line).
- Parse the expression (adding 7 to the col_offset of each parsed node, lineno remains the same since it's 0).
- Fix the node locations by shifting the Name node by 14, which is the number of characters in the f-string (counting the `f` and the opening quote) before the start of the expression.

With this change we now pre-feed the parser with the exact lineno and offset of the expression itself, not the f-string. This allows us to completely skip the third step of shifting the node locations.