Lazy load ML dependencies

[donatasm]

Bringing back lazy loads reverted in #356, because of pip not supporting publishing packages referencing git commit hashes.

Support for Python 3.7 was added in scientific-python/lazy-loader#14, hence package is referenced by version now.

Not bringing back lazy loading pandas or numpy as they're always are loaded anyway, because of the image extensions.

Closes:

• https://linear.app/layer/issue/LAY-3689/lazilly-load-sdk-dependencies-that-are-not-always-needed
• https://linear.app/layer/issue/LAY-3341/import-kerascallback-and-xgboostcallback-lazily

[linear]

LAY-3689 Lazilly load sdk dependencies that are not always needed

Description

part of https://github.com/layerai/roadmap/issues/19

Currently import layer is quite slow. We should reduce the number of large dependencies we load automatically and add a test to ensure we don't regress this if possible.

An inspiration can be taken in what we did last time we wanted to speed up the SDK - https://github.com/layerai/monorepo/pull/3146/files

Acceptance Criteria

• [ ]

[cihati]

nit: I'd probably assert something expected to be in deps as well, to make this more sturdy.