Drop pandas on DBR 16.4 + serverless-v3 (no cp312 wheel)

[rugpanov]

Stacked on #21 (base branch = envgen/drop-nonlocal-constraints). It builds on that PR's envgen.py and regenerated artifacts, so it should merge after #21; I'll retarget to main once #21 lands.

Closes #18.

Why

DBR 16.4 and serverless-v3 pin pandas~=1.5.3 under requires-python==3.12.* — faithful to the image (confirmed against the DBR 16.4 LTS release notes, which list pandas 1.5.3 on Python 3.12.3). But pandas 1.5.3 has no cp312 wheel — pandas ships Python 3.12 wheels only from 2.1.1 — so uv sync / pip install -c can't install it locally and falls back to a failing source build.

These are the only two Python-3.12 runtimes still on pandas 1.5.x: 13.3/14.3/15.4 predate 3.12, and 17.3+ / serverless-v4+ already ship pandas 2.x. No in-range (~=1.5) version has a cp312 wheel, so widening can't salvage it.

What

• envgen.py — add DROP_BY_ENV, an environment-scoped drop applied in _filtered (threaded env_name through build_pyproject / build_constraints). pandas is dropped for 16.4.x-scala2.12, 16.4.x-{cpu,gpu}-ml-scala2.12, and serverless-v3 only; every other env keeps its pandas pin. Documented in the module docstring, the DROP_BY_ENV comment, and the README "What is intentionally not included" section.
• test_envgen.py — EnvScopedDropTest covers the drop on the four targets and the keep elsewhere.
• Regenerated the four target envs' pyproject.toml + constraints.txt (pandas removed).

Result: pandas resolves to an installable version locally on those two runtimes, instead of a hard uv sync failure. (The durable general form — a PyPI wheel-availability guard in sync.py — is noted as future work on #18.)

Verification

• python -m unittest test_envgen — 13 passing.
• pandas absent from the 4 targets, present in 15.4 (1.5.3), 17.3 (2.2.3), serverless-v4 (2.2.3); all 33 pyproject.toml valid TOML.

This pull request and its description were written by Isaac.