perf(contracts): optimize contract gas usage and establish benchmarks

[mftee]

Summary

• counter: COUNTER moved from persistent to temporary storage. Temporary entries are cheaper to read/write (no long-term rent accounting). Safe here because the counter carries no financial value — worst case on TTL lapse is a silent reset to 0.
• betting: removed MatchStats.bettors: Vec<Address> — it was appended to and the entire MatchStats record rewritten on every single place_bet call, but the field was never read anywhere (not even by claim_payout). This made place_bet's storage cost grow linearly with the number of bettors already on a match. Removing it restores constant per-call cost.
• betting: get_bet/Bet intentionally kept on persistent storage, with the rationale documented in the module doc comment — Bet holds real transferred funds, a claimed flag, and a payout; moving it to temporary storage risks unrecoverable fund loss if the TTL lapses before a bettor claims/refunds. The issue asked to move get_count/get_bet to temporary storage "where safe" — for get_bet specifically, it isn't.
• Added gas_bench test modules to counter and betting that measure cpu_instruction_cost/memory_bytes_cost per function via env.budget() and assert against configurable thresholds (see docs/gas-benchmarks.md).
• Added a Gas budget benchmarks step to .github/workflows/contracts.yml running these two gas_bench suites so a regression fails CI.

A blocking pre-existing issue I found and partially fixed

None of counter, betting, rewards, or player-nft's Cargo.toml declared soroban-sdk's testutils feature in [dev-dependencies] (oracle and vault already had the correct pattern). That meant Env::mock_all_auths, Address::generate, env.register_contract, etc. were unavailable, and none of these four crates' #[cfg(test)] modules had ever actually compiled — invisible until now because contracts.yml only ran cargo check, never cargo test. I fixed the Cargo.toml gap in all four (copying the working pattern from oracle/vault), since my new gas_bench tests need cargo test to actually work.

This surfaced further pre-existing, unrelated bugs in betting's existing mod test that were also never caught for the same reason (several call sites treat an already-auto-unwrapped client method return value as a Result and call .unwrap() on it again; one helper function is missing a lifetime annotation). I fixed two of these opportunistically while getting my own code to compile (the upgrade/try_upgrade call in test_upgrade_requires_two_governance_approvals_and_preserves_state, and the missing lifetime on the initialize test helper), but did not do a full audit of the rest of mod test — that's a pre-existing problem spanning the whole file and out of scope for this gas-optimization issue.

Practical consequence: I could not get a clean local cargo test --release run to completion — beyond the remaining pre-existing test bugs, this Windows dev environment also hit unrelated toolchain issues (missing dlltool.exe for the -gnu target, incomplete MSVC "C++ build tools" for the -msvc target) that block linking locally regardless. The Gas budget benchmarks CI job on this PR is the first environment where this can actually be verified end-to-end, and it may currently fail on the remaining pre-existing bugs in betting's mod test rather than on anything in this diff. I'd rather flag that clearly than claim a green run I couldn't produce. Happy to keep iterating from the actual CI output, or to split the pre-existing-bug cleanup into its own PR if that's preferred.

docs/gas-benchmarks.md

Records the budget thresholds enforced by gas_bench and the rationale for each storage decision above. The "Actual" measured values are marked pending the first real CI run for the reason above, rather than backfilled with numbers I couldn't actually produce locally.

Closes #23