Auto merge of #98748 - saethlin:optimize-bufreader, r=Mark-Simulacrum

Remove some redundant checks from BufReader

The implementation of BufReader contains a lot of redundant checks. While any one of these checks is not particularly expensive to execute, especially when taken together they dramatically inhibit LLVM's ability to make subsequent optimizations by confusing data flow increasing the code size of anything that uses BufReader.

In particular, these changes have a ~2x increase on the benchmark that this adds a `black_box` to. I'm adding that `black_box` here just in case LLVM gets clever enough to remove the reads entirely. Right now it can't, but these optimizations are really setting it up to do so.

We get this optimization by factoring all the actual buffer management and bounds-checking logic into a new module inside `bufreader` with a new `Buffer` type. This makes it much easier to ensure that we have correctly encapsulated the management of the region of the buffer that we have read bytes into, and it lets us provide a new faster way to do small reads. `Buffer::consume_with` lets a caller do a read from the buffer with a single bounds check, instead of the double-check that's required to use `buffer` + `consume`.

Unfortunately I'm not aware of a lot of open-source usage of `BufReader` in perf-critical environments. Some time ago I tweaked this code because I saw `BufReader` in a profile at work, and I contributed some benchmarks to the `bincode` crate which exercise `BufReader::buffer`. These changes appear to help those benchmarks at little, but all these sorts of benchmarks are kind of fragile so I'm wary of quoting anything specific.