Modify pyproject.toml and make mamba_ssm optional.

[giogix2]

What has changed?

• Made the libraries mamba-ssm and causal-conv1d optional. Now they can be installed by specifying [mamba-ssm] within square brackets
• Updated README.me to reflect the changes.
• Removed setup.py.

What was causing the issue?

Originally the library mamba-ssm was using setup.py to install the package. However, this library expected torch to be already installed (see here), which setup.py was importing at the top of the file. This is why a fresh installation of mamba-ssm would fail.

How was it resolved.

The installation process was improved by introducing pyproject.toml as part of this PR, which was mentioned in this GitHub issue.

NOTE

Currently the new version (v2.2.3) of mamba-ssm is not available in Pypi yet (link). As of today (2024/12/06) the CI/CD to release the new package is running (though unsuccessfully). Until is available, this solution installs the packages from github, specifying the tag v2.2.3 to fix the version.

[bputzeys]

I think we should test this in a fresh environment. I suggest to do this on the release.yml.

• do not run on self hosted to get a clean environemnt https://github.com/helicalAI/helical/blob/release/.github/workflows/release.yml#L13
• pip install .[mamba-ssm] here: https://github.com/helicalAI/helical/blob/release/.github/workflows/release.yml#L26
• we have to check if all the tests can be run on a non self-hosted environment which does not have GPU.

[giogix2]

FYI for reviewers. I'm going to add new non self-hosted runners on another branch/PR. I'll update this one when that's done.