docs: follow-ups on the HTTPS outcalls pricing pages

[marc0olo]

Summary

Follow-ups on #254. Wording and one table; no behavior change.

references/cycle-costs.md

• The version 2 table carried only the base and delivery terms, so the usage fee was invisible to anyone pricing a call from it. Added rows for downloaded bytes, round-trip time and transform instructions, labeled each row with the part of the price it belongs to, and said which nodes pay them.
• Said what the 13 dividing transform_instructions is: the reference subnet size the fees are calibrated against, not the node count.

guides/backends/https-outcalls.mdx

• The PUT/DELETE/PATCH conditions move from the opening paragraph into a row of the modes table.
• "Testing locally" now says that flexible outcalls do run locally, and what a local run does not exercise.
• "Depends on the CDK" now names the wrapper: Motoko is a language purpose-built for ICP, not a CDK.
• Added send_http_flexible to the upstream provenance comment.

concepts/https-outcalls.md

• Replaced the paragraph that named the Rust builders and repeated the guide's Call.httpRequest warning with a pointer to the guide, which is where someone writing the call reads it.

Verification

• Every figure in the version 2 table was re-derived from rs/https_outcalls/pricing/src/fees.rs at 4be965e1 (the replica elected by proposal 143907) and cross-checked against the six totals that file's own tests assert. The 34-node column is the same formulas at n = 34: http_request_fee_v2 adds no subnet-size scaling of its own. The rows are what a call settles at, not what ic0.cost_http_request_v2 reserves.
• The method conditions in the modes table are the two validation sites in rs/types/types/src/canister_http.rs (DeterministicResponseCountRequired).
• send_http_flexible was run on a local network with icp-cli 1.5.0 and returned all five responses it asked for.
• npm run build passes.

[marc0olo]

Feedback addressed:

• Round-trip time does not dominate. You are right, and by more than the comment claims: at 13 nodes round trip is capped at 60 s = 234M cycles, while a transform at the full 5B instruction limit costs ~5B and delivering a 2MB response ~19B. One delivered byte already costs 2.43 ms of round trip, so delivery outweighs round trip about tenfold even on a 1 KB, 500 ms call. Replaced the claim with the three ceilings.
• The local node count. Dropped it. icp-cli's local network is PocketIC: one process, but its registry declares the application subnet with 13 nodes (subnet_size(Application) in rs/pocket_ic_server/src/pocket_ic.rs), so committee validation and cycle costs behave as on a 13-node subnet while that one process makes every request. Stating that next to "the local replica runs a single node" read as a contradiction, and it is local-tooling internals. The note now keeps only the part that affects a test: responses cannot disagree locally.