feat(a16): real in-session benchmark capture and compare

[Gift-Stack]

bench.listSnapshots and bench.compare were empty stubs — and the contract had no way to even create a run. A16 now has a real, self-contained benchmark engine built from the sandbox's own executed-transaction metrics (no dependency on @ton/sandbox/jest-reporter files, which need a Blueprint/Jest project).

What this does

• New bench.capture method (contract + client + hook + both mocks in lockstep): snapshots real aggregates from the workspace's tx history — transactions, total/compute/action/storage fees, compute gas, message cells/bits, max tree depth. All money/gas as bigint→decimal strings.
• bench.listSnapshots returns stored runs (newest first); bench.compare computes real per-metric diffs — deltaPct guarded for zero baselines, regression flagged only when a lower-is-better metric grew.
• BenchmarkPanel gets a capture bar; existing run-vs-baseline pickers + delta table now work against real data.
• Fixed a field-name bug found during verification: action phase message size is totalMessageSize in @ton/core 0.63 (totalMsgSize doesn't exist), which had cells/bits reading 0.

Verification

• pnpm tsc green; bridge smoke ALL OK.
• Live check (1 send → baseline; +1y, 3 sends → current): all 9 metrics real and non-zero — e.g. total fees 280757→1903107 (+577.84%, regression), compute gas 2246→8984 (+300%, regression), storage fees 0→779546 (regression), tree depth 2→2 (0%, not flagged). deltaPct math and regression flags script-asserted.