[doc-only] docs: document setuptools-scm clone requirements for source builds

[bharatr21]

Description

Closes #2400

Every package in this repo derives its version from git tags via setuptools-scm,
but there's no mention in the docs, and the git clone commands in our own install guides
that are intended to be copied by contributors produce broken builds.

Two of the three bad-clone cases fail silently:

Clone	Resolved version
--depth 1	0.1.dev1+g0d22cb444 — no error
--no-tags	0.1.dev2114+g0d22cb444 — no error
source zip (no .git)	LookupError — the only case that errors

A partial --depth is worse than none. With the nearest cuda-pathfinder-v* tag
10 commits back, a --depth 20 clone gives:

Package	Version	Outcome
cuda_pathfinder	1.6.1.dev10+g0d22cb4	Correct, tag falls inside depth
cuda_core	0.1.dev20+g0d22cb444	Wrong, no error
cuda_bindings	0.1.dev20+g0d22cb444	Wrong, no error

So "it built and the version looked right" is not evidence the clone is sound,
and no fixed --depth stays correct as commits land between releases.

Changes

• CONTRIBUTING.md: new "Cloning the repository" section (plus TOC entries)
  covering the per-package tag patterns, why root = ".." means the whole repo is
  needed, the recommended treeless clone, recovery for shallow clones and stale
  forks, and the symptoms of each failure mode.
• Install guides: updated the four git clone commands in the cuda_core and
  cuda_pathfinder guides to git clone --filter=blob:none, and added a pointer to
  the new section from all three. cuda_bindings has no clone command, so it gets a
  third requirement bullet alongside the existing CTK header/static-lib ones.

The recommended clone matches what CI already does (fetch-depth: 0 +
filter: blob:none in build-wheel.yml): full commit graph and tags, no historical
blobs. It took 6.2s against github.com.

Verification

• Treeless clone from GitHub fetches all 87 tags, is not shallow, and has the full
  2114-commit graph; all four packages resolve correct versions
  (13.3.2.dev128, 1.1.1.dev32, 1.6.1.dev10) via python -m setuptools_scm,
  so each package's real pyproject.toml config is exercised.
• --depth 1, --depth 20, --no-tags, and no-.git failure modes all reproduced
  against the real remote.
• SETUPTOOLS_SCM_PRETEND_VERSION_FOR_CUDA_CORE and git fetch --unshallow --tags
  recovery paths confirmed working before documenting them.
• Full pre-commit hook set passes on all four files with no reformatting; RST
  parses clean and lychee reports no broken links.

Checklist

• I am familiar with the Contributing Guidelines.
• New or existing tests cover these changes.
• The documentation is up to date with these changes.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[mdboom]

The --filter=blob:none is not necessary -- it just reduces the size of the clone while still making it buildable. It's perfectly fine to just use the "default" clone behavior of git, so I think --filter=blob:none just confuses things and should not be mentioned.

If we are concerned about repository size, we should periodically flatten the history of the gh-pages branch. (That is something we would do in automation -- no need for most developers to know about it).

[mdboom]

This is way too much detail.

Suggested change

	Each package is tagged on its own cadence, so the distance from `main` back to
	the relevant tag differs per package and grows as commits land between releases.
	There is no `--depth` value that is safe for all four packages, and any value
	that works today will silently stop working later.
	This makes a partial `--depth` actively dangerous, because it can succeed for one
	package while quietly failing for the others. A `--depth 20` clone taken while
	the nearest `cuda-pathfinder-v*` tag was 10 commits back gives:
	```text
	cuda_pathfinder -> 1.6.1.dev10+g0d22cb4 # correct
	cuda_core -> 0.1.dev20+g0d22cb444 # wrong, no error
	cuda_bindings -> 0.1.dev20+g0d22cb444 # wrong, no error
	```

[mdboom]

I think we just recommend the defaults, tbh.

Suggested change

	### Recommended clone
	```console
	$ git clone --filter=blob:none https://github.com/NVIDIA/cuda-python.git
	```
	This is a *treeless* clone: it fetches the complete commit graph and all tags
	(which is what `setuptools-scm` needs) while skipping historical file contents
	(which it does not). It is substantially faster than a plain clone and is
	exactly what CI uses — see the `fetch-depth: 0` plus `filter: blob:none`
	checkout settings in `.github/workflows/build-wheel.yml`.
	Do **not** use `--depth`, `--shallow-since`, or `--no-tags`.