Add API-backed dataset cost estimates

[dianaprior]

Change Description

Adds estimate_cost(X_train, X_test, ...) so callers can check an operation's quota cost before running it. The helper sends only dataset dimensions and settings to POST /tabpfn/estimate_cost, returning a typed result with estimated_cost, pricing_version, and the resolved inputs. It does not upload feature values or consume quota.

Supports prediction, Thinking fit/predict, and cached prediction. Defaults are resolved by the server; cache fallback or different fitted estimator counts can change the final charge. The endpoint requires quota v3 to be enabled.

Server counterpart: https://github.com/PriorLabs/tabpfn-server/pull/1164 (endpoint commit 78435846).

from tabpfn_client import estimate_cost

quote = estimate_cost(X_train, X_test, model_version="v3")
print(quote.estimated_cost, quote.pricing_version)

Validation: all 275 client unit tests passed, including 17 new helper tests. Ruff, type checks, and direct client-to-server HTTP contract checks also passed locally.

New dependencies: None.

Mattermost reviewer ping: None; this PR is a draft.

Breaking changes

Does this PR break the API? No. The new helper requires the server counterpart above; existing client calls are unchanged.

Does this PR break the user interface? No.

[CLAassistant]

All committers have signed the CLA.

[safaricd]

All clean - thanks!