Support custom MAXTEXT_DIR and --quantization option

[jberchtold-nvidia]

This PR updates test-maxtext.sh with two features.

1. If a user specifies MAXTEXT_DIR in their environment, this dir will be used instead of /opt/maxtext. This makes it easier for development. Default behavior of using /opt/maxtext is unchanged unless that environment variable is specified.
2. Support for more quantization recipes via MaxText's quantization config via a new --quantization arg.

Before this PR, this script uses --dtype=fp8 to specify MaxText FP8 DelayedScaling. However, when developing new quantization recipes, support for different types of quantization beyond "fp8" is useful. For backwards compatibility, I've kept the behavior of --dtype=fp8 unchanged. But you can now specify --quantization="..." with an arbitrary string and it will be passed to MaxText's quantization config unfiltered. An error is raised when both --dtype=fp8 and --quantization="..." are specified at the same time.

[jberchtold-nvidia]

Closed and replaced with #1622 which uses a branch on the JAX-Toolbox repo instead of the fork