This is a pretty major, but surprisingly not that invasive, overhaul of
how binaries
are started. It fixes several issues and lays ground work for future
improvements.

In brief:

* A system Python is no longer needed to perform bootstrapping.
* Errors due to `PYTHONPATH` exceeding environment variable size limits
is no
  longer an issue.
* Coverage integration is now cleaner and more direct.
* The zipapp `__main__.py` entry point generation is separate from the
Bazel
  binary bootstrap generation.
* Self-executable zips now have actual bootstrap logic.

The way all of this is accomplished is using a two stage bootstrap
process. The first
stage is responsible for locating the interpreter, and the second stage
is responsible
for configuring the runtime environment (e.g. import paths). This allows
the first
stage to be relatively simple (basically find a file in runfiles), so
implementing it
in cross-platform shell is feasible. The second stage, because it's
running under the
desired interpreter, can then do things like setting up import paths,
and use the
`runpy` module to call the program's real main.

This also fixes the issue of long `PYTHONPATH` environment variables
causing an error.
Instead of passing the import paths using an environment variable, they
are embedded
into the second stage bootstrap, which can then add them to sys.path.

This also switches from running coverage as a subprocess to using its
APIs directly.
This is possible because of the second stage bootstrap, which can rely
on
`import coverage` occurring in the correct environment.

This new bootstrap method is disabled by default. It can be enabled by
setting
`--@rules_python//python/config_settings:bootstrap_impl=two_stage`. Once
the new APIs
are released, a subsequent release will make it the default. This is to
allow easier
upgrades for people defining their own toolchains.

The two-stage bootstrap ignores errors during lcov report generation,
which
partially addresses
#1434

Fixes #691

* Also fixes some doc cross references.
* Also fixes the autodetecting toolchain and directs our alias to it

This PR will reduce the time it take to run `bazel coverage` 

*Before this fix,*

 ```
 bazel coverage  --cache_test_results=no  //...
 Elapsed time: 7.054s, Critical Path: 6.87s
 
 ```
 
[lcov --list
bazel-out/_coverage/_coverage_report.dat](https://github.com/user-attachments/files/16681828/lcov.log)


*After*
```
 bazel coverage  --cache_test_results=no  //...
Elapsed time: 2.474s, Critical Path: 1.18s


$ lcov --list bazel-out/_coverage/_coverage_report.dat 
Reading tracefile bazel-out/_coverage/_coverage_report.dat
                                          |Lines       |Functions  |Branches    
Filename                                  |Rate     Num|Rate    Num|Rate     Num
================================================================================
[/home/ewianda/.cache/bazel/_bazel_ewianda/da4b4cc49e0e621570c9e24d6f1eab95/execroot/_main/bazel-out/k8-fastbuild/bin/benchsci/ml/nlp/]
test_tokenizer_stage2_bootstrap.py        | 6.0%    250|    -     0|    -      0

[benchsci/]
devtools/python/pytest_helper.py          |90.5%     21|    -     0|    -      0
ml/nlp/test_tokenizer.py                  | 100%     13|    -     0|    -      0
ml/nlp/tokenizer.py                       |61.8%     76|    -     0|    -      0
================================================================================
                                    Total:|26.1%    360|    -     0|    -      0

```

Related to #1434

---------

Co-authored-by: aignas <[email protected]>