sdk/shreds: bound RPC client request timeout and connection pool

[nikw9944]

Summary of Changes

• Configure the shreds SDK RPC client with a bounded 15s per-request timeout and a sized connection pool (128) instead of wrapping the unbounded, lightly-pooled http.DefaultClient.

This is the SDK-level counterpart to the internet-latency-collector hardening (#3973). http.DefaultClient has no request timeout and keeps only 2 idle connections per host, so against a slow or degraded ledger RPC endpoint a request can block indefinitely — long enough for a transaction's recent blockhash to expire before it is sent, which surfaces as BlockhashNotFound at preflight. A bounded timeout makes requests fail fast so the caller can retry with a fresh blockhash. The existing transient-error retry wrapper is unchanged.

Diff Breakdown

Category	Files	Lines (+/-)	Net
Scaffolding	1	+34 / -3	+31
Tests	1	+43 / -0	+43
Docs	1	+2 / -0	+2
Total	3	+79 / -3	+76

Small change: RPC-client construction hardening plus a test; no logic/algorithm changes.

Key files (click to expand)

• sdk/shreds/go/rpc.go — add newHTTPClient (bounded timeout + sized transport) and use it for the RPC client's inner HTTP client instead of http.DefaultClient.
• sdk/shreds/go/rpc_test.go — new tests asserting the client has a bounded timeout and that a slow request is aborted near the timeout rather than hanging.

Testing Verification

• TestNewHTTPClient_HasBoundedTimeout: the constructed client has a non-zero timeout and a connection pool sized to defaultMaxConns (guards against regressing back to http.DefaultClient).
• TestNewHTTPClient_TimesOutSlowRequest: a request to a handler that sleeps 3s is aborted within ~200ms, confirming requests are bounded.

Note

The identical unbounded-http.DefaultClient pattern exists in sdk/revdist/go/rpc.go; left out of this PR to keep it focused, but worth a follow-up.