Budgets exist but meter nothing for self-hosted models

`/v1/projects//budget` returns `max_budget_usd`, `spend_usd` and
`remaining_usd`, and `/v1/projects//usage` returns entries with
`total_cost_usd`. The plumbing works.

But **`spend_usd` reads 0.00 regardless of usage.** Self-hosted models have no
entry in LiteLLM's cost map, so every completion is priced at zero. LiteLLM
warns about this on startup:

```
register_model: model=openai/gemma4-26b not in built-in cost map and no
prefix/region variant matched; cache cost fields will default to 0
```

So today the endpoints give **request attribution, not dollars** — and a
project budget can never be exhausted because nothing accrues against it.

Three things to decide:

**What does a self-hosted token cost?** Options: amortised GPU rental over
measured throughput (the A100 is ~$1.19/hr and does ~400 tok/s aggregate, so
roughly $0.0000008/token), a nominal figure for fair-share accounting, or
genuinely zero with budgets used only for rate-limiting.

**Where does it go?** `input_cost_per_token` / `output_cost_per_token` in the
route's `model_info`, set when llmao registers the route. That makes it part of
the same push as the fleet work rather than a separate mechanism.

**What happens at the limit?** LiteLLM blocks the key when a team exceeds its
budget. Worth confirming the failure is legible — a 429 with a clear message
rather than something that reads as an outage.

Note the reporting is only as good as the price. A wrong-but-nonzero figure is
more useful than zero for spotting a runaway job, and less useful than nothing
if anyone treats it as real money. Worth labelling in the UX either way.

Start with the /v1/projects//budget and /v1/projects//usage behavior, then trace how LiteLLM receives input_cost_per_token and output_cost_per_token in the route's model_info. Define a pricing policy for self-hosted tokens and verify that spend accrues, budgets can be exhausted, and the limit response is a clear 429 rather than an outage-like failure.

Written by the indexing model from the issue text.