Add Chinchilla uncertainty only if an identifiable N,D grid exists

[esaran1]

Why

--fit-form chinchilla can point-project when params_n and tokens_d are present. Gate uncertainty and projection bootstrap still raise NotImplementedError for that form. A one-dimensional compute fit can also look like saturation when the true dependence is separate in N and D.

What to do

Only start this if the real grid varies N and D independently. Then:

1. Add a stratified bootstrap (or another stated method) for Chinchilla projections.
2. Fail loudly on collinear log N, log D designs; do not fall back to compute power laws.
3. Project to a stated target (N, D), not by reading those values from the target BPB rows if the target design is supposed to be chosen in advance.
4. Tests: recover known two-term parameters; reject coupled D=N^k designs; reject missing columns.

Acceptance

• Uncertainty exists or the CLI still refuses gate/CIs for Chinchilla rather than pretending
• No silent fallback to compute_power_law
• Identifiability checks are tested

Files

• asla/models.py
• asla/analysis/fits.py
• asla/analysis/gate.py
• tests/test_chinchilla.py

Do not fabricate

• Invent N,D for compute-only runs
• Enable Chinchilla on the placeholder 45-run relative-compute grid unless those columns are actually logged