feat(catalog): add optional tiered-price and cache-write price fields

[argszero]

Summary

Adds the expression capability and validation for two price features the catalogue
currently cannot express (rant 2026-10-16T14:08:14, 【计价】tiered input price +
cache-write price):

1. Per-request input-length tiers — an optional [[models.tiers]] array:

   [[models.tiers]]
   from_tokens = 512001
   input_per_m = 4.20
   cache_hit_input_per_m = 0.84
   output_per_m = 16.80

   from_tokens is the inclusive lower bound of the tier; the model's own
   input_per_m / cache_hit_input_per_m / output_per_m are the first tier.

2. Cache-write price — an optional cache_write_input_per_m field, defaulting to
   0 ("cache writes are free", i.e. today's behaviour).

Both fields are #[serde(default)], so every existing catalogue row parses exactly as
before and pricing is byte-for-byte unchanged — settle is deliberately not touched
in this step.

Config::validate gains the rules that the data itself cannot be trusted to satisfy:
tier from_tokens must be > 0 and strictly increasing, and no tier price (or
cache_write_input_per_m) may be negative. Tier data goes straight into real money, so
a malformed catalogue is rejected at load time instead of mis-pricing silently.

Scope — this is step 1 of 2

The rant asks for two steps; this PR is the first:

• Step 1 (this PR): the fields exist, parse and validate. Nothing in the pricing
  path reads them yet.
• Step 2 (follow-up): wire them into billing — propagate the tier data through
  src/db.rs::seed_models into the models table and select the tier by the request's
  input-token count in settle, and per-tier boundary tests for each vendor.

The tests below pin the backward-compatibility half of the rant's acceptance criteria
(no-tier rows price identically — settle and its expectations are untouched, suite
unchanged counts).

Related Issue

Rant 2026-10-16T14:08:14 (【计价】按输入长度分档的价格、以及缓存写入价，都没有字段).

Tests

• Existing tests pass — cargo test: 571 passed / 0 failed (564 before; the 7
  new tests are below)
• New tests added:
  • parse_model_tier_from_toml — a tier parses field-by-field; cache_hit_input_per_m defaults to 0
  • model_defaults_to_no_tiers_and_zero_cache_write — every example row defaults to empty tiers / zero cache-write price
  • validate_accepts_strictly_increasing_tiers — positive control
  • validate_rejects_non_increasing_tier_boundary — duplicate and decreasing boundaries
  • validate_rejects_zero_tier_boundary
  • validate_rejects_negative_tier_price
  • validate_rejects_negative_cache_write_price
• cargo fmt clean
• cargo clippy --all-targets -- -D warnings clean
• non-test build clean (RUSTFLAGS="-D warnings" cargo build) — the only witness
  that the new fields are read by production code, not just tests