simplify speedups implementation

[davidism]

Only implement the actual string replacement in C. escape, escape_silent, and soft_str are always pure Python. This greatly simplifies the C code, as it no longer needs to do type checks, __html__ check, or string conversion. No longer requires a circular import between Markup and escape. The tests also become simpler, not requiring fixtures for each function. Moving all that to Python only slows down the benchmark by about 20 nanoseconds, which I find acceptable compared to the simpler code.

@carsonburr did this when working on a Rust implementation, as it reduced the amount of non-Rust work done in the Rust code.