Fix malformed versions from push-triggered builds on release branches

[AyodeAwe]

Summary

• Fix rapids-generate-version to handle the case where the nearest git tag has no stage (e.g., v25.10.00 instead of v25.10.00a)
• When {stage} is empty, dunamai's {base}{stage}{distance} format produces garbage like 25.10.000 or 26.02.010
• The fix detects an empty stage and defaults to a as the separator

Context

During the cucim v26.02.01 hotfix (ref: ops#4293), the push to release/26.02 triggered a build that produced packages with malformed versions (libcucim-26.02.010). The tag-triggered build produced the correct version (26.02.01).

This happens because push-triggered builds use {base}{stage}{distance} but when the nearest tag is a release/hotfix tag, {stage} is empty -- so distance concatenates directly onto base.

Verification

Tested locally against cucim (the repo where the bug was originally reported) and rmm:

cucim:

Scenario	Tag	Before	After
Hotfix push (the original bug)	v26.02.01	26.02.010	26.02.01a0
Release push	v26.02.00	26.02.000	26.02.00a0
Nightly on main	v26.04.00a +18	26.04.00a18	26.04.00a18

rmm:

Scenario	Tag	Before	After
Release push	v26.02.00	26.02.000	26.02.00a0
Release push	v25.12.00	25.12.000	25.12.00a0
Nightly on main	v26.04.00a +52	26.04.00a52	26.04.00a52

How to reproduce

# Test against cucim (where the bug was originally reported)
git clone git@github.com:rapidsai/cucim.git
cd cucim
pip install dunamai

# Bug: at the hotfix tag, {stage} is empty so distance concatenates onto base
git checkout v26.02.01
python -m dunamai from git --format "{base}{stage}{distance}"
# Output: 26.02.010 (wrong -- this is what James reported)

# Bug: same issue at a release tag
git checkout v26.02.00
python -m dunamai from git --format "{base}{stage}{distance}"
# Output: 26.02.000 (wrong)

# Fix: detect empty stage and default to "a"
git checkout v26.02.01
python -c "
from dunamai import Version
v = Version.from_git()
if v.stage:
    print(v.serialize(format='{base}{stage}{distance}'))
else:
    print(v.serialize(format='{base}a{distance}'))
"
# Output: 26.02.01a0 (correct)

[jameslamb]

Thanks so much for investigating this! The description and comments helped me a lot.

I think applying a fix here is a good approach. I'm unsure if we need that one more nightly package pushed to the nightly indices, but it's not too expensive to have it (since CPU CI runners are relatively cheap and plentiful).

But I left one suggestion for a slightly different version for those.

[AyodeAwe]

Thanks @jameslamb! Addressed!

[jameslamb]

Thanks for considering it, let's merge this