fix(data-store): latch bytes-unsupported instead of re-probing every sync

[vaidehi-figma]

Problem

DataStoreTrait::get_bytes / set_bytes have default impls that return BytesNotImplemented, and whether an implementation overrides them is a static property of that implementation. The C FFI shim (statsig-ffi/src/data_store_c.rs) exposes no bytes hook at all, so every probe from a Go/C consumer refuses — forever.

Both the read and write paths treated that refusal as a fresh event, so a string-only data store logs a WARN on every specs sync:

WARN [Statsig::StatsigDataStoreSpecsAdapter] Failed to read specs from data store as bytes. Falling back to string read: Bytes method not implemented
WARN [Statsig::SpecStore] Data store bytes write is not implemented. Falling back to string write: Bytes method not implemented

At the default 10s specs_sync_interval_ms that's 12 warnings/minute/process, forever, plus two doomed FFI round-trips per sync. Figma's labmate runs as a daemonset with an S3-backed data store, so this currently dominates its log volume in staging and production.

Nothing is functionally broken — the string fallback works, flags evaluate correctly. It's noise plus a little wasted work.

Fix

Latch the first refusal in both places.

Read path — StatsigDataStoreSpecsAdapter gains a bytes_unsupported: AtomicBool. Once set, load_cached_specs goes straight to the plain-text string read instead of re-probing get_bytes. The fallback warning is emitted on the first refusal only.

Write path — SpecStore gains a DataStoreBytesWriteSupport latch, threaded into write_specs_to_data_store, which skips set_bytes once it's known to refuse. The string-fallback and protobuf-skip warnings get separate "already logged" flags, so a protobuf response still reports its dropped write even if a plain-text response already reported the string fallback — that one means specs aren't cached at all and shouldn't be swallowed.

Behavior is otherwise unchanged: same keys read and written, and genuine (non-BytesNotImplemented) data store errors still log on every occurrence.

Testing

• New test_string_only_data_store_probes_bytes_once_across_syncs asserts a string-only store is probed for bytes exactly once across many background syncs. It reports 4 probes without the fix, 1 with it.
• cargo test --test data_store_bytes_adapter_tests --test proto_specs_usage_tests --test customized_specs_adapter_tests — all pass.
• cargo clippy --all-targets clean; touched files are rustfmt-clean.

Follow-ups (not in this PR)

• Cut a figma3 release so labmate can pick this up (both statsig-go and binaries-linux-gnu tags), then bump the two replace directives in the figma monorepo go.mod.
• Worth sending upstream: statsig-io/statsig-server-core 0.22.0 has the same per-sync log_w! and the same bytes-less C FFI.
• The real long-term fix is a bytes hook in data_store_c.rs + the Go binding, which would let C/Go consumers use the statsig-br cache instead of plain text.