perf: Speed up compile times with nom

[epage]

This is using a short-lived fork nom until we can get the changes upstreamed. Note that combine requires using attempt to backtrack while nom backtracks by default and requires cut to not backtrack.

I find the straight functions much easier to understand what is happening and this allows us to intermix procedural with declarative logic (e.g. in string processing I skipped combinators to make it easier to not require an allocation).

Regarding that allocation, we still do it but this opens the door for us to use InternalString for user values which might give us more of a performance boost (previously, the forced allocation made it a moot point to measure it).

Running cargo clean -p toml_edit && time cargo check, I get 3s for building toml_edit with combine and 0.5s for nom.

For runtime performance

• Parsing cargo inits generated manifest took 4% less time
• Parsing cargos manifest took 2% less time
• 10 tables took 37% less time
• 100 tables took 41% less time
• Array of 10 tables took 38% less time
• Array of 100 tables took 40% less time

(some of this might be because we use dispatch more than before and some because our error reporting is less allocation heavy)

This is with Rust 1.66 on a i9-12900H processor under WSL2

Fixes #327

[epage]

Also, compare CI times

• Before: https://github.com/toml-rs/toml/actions/runs/3754649195
• After: https://github.com/toml-rs/toml/actions/runs/3767930124