Run the test suites against managed regtest backends instead of docker compose

[bennyhodl]

Run the test suites against managed regtest backends instead of docker compose

cargo test now needs nothing running beforehand. Each test binary starts and
tears down its own bitcoind, electrs (serving the esplora HTTP API), PostgreSQL
server, and nostr relay on ephemeral ports.

Before this, every integration test read its connection details out of the
environment (BITCOIND_HOST, ESPLORA_HOST, DATABASE_URL, ...) and expected
the docker-compose stack to already be up on fixed ports. CI reproduced that with
docker compose up -d plus a pair of polling scripts in front of every test job.

What replaces what

docker-compose service	replacement
bitcoin + electrs	bitcoind 0.41 + electrsd 0.41 (esplora_a33e97e1)
postgres	postgresql_embedded 0.19
nostr-relay	nostr-relay-builder 0.44 (in-process)

docker-compose.yaml stays: it is still how you run a ddk-node locally
(just deps). It is simply no longer involved in testing.

The new testenv crate

ddk-testenv is a workspace member with publish = false, pulled in by ddk
and ddk-manager as a path-only dev-dependency. Cargo strips such dependencies
when packaging, so releases are unaffected — verified with cargo package.

It offers two ways to get backends, and the distinction matters:

• ddk_testenv::env() — one bitcoind/electrs pair per test binary, shared by
  every test in it. Cheap, and fine for tests that just need somewhere to put
  coins.
• TestEnv::new() — a private pair, torn down when dropped. Tests that assert on
  contract state as blocks advance need this. libtest runs tests concurrently,
  and on a shared chain the blocks a sibling test mines can push a contract past
  a locktime before the assertion runs. The manager execution and splice tests
  use it, which is why init_clients and create_and_fund_wallet gained an
  &TestEnv parameter.

Mining helpers now block on the indexer (ElectrsD::wait_height) rather than
sleeping. That replaces the sleep(Duration::from_secs(5)) polling loops the old
helpers used, and is where most of the wall-clock saving comes from.

Details worth reviewing

Process cleanup. libtest exits via std::process::exit, which never runs
destructors for statics, so the shared environment's child processes would
outlive the test binary as orphans. env() registers a libc::atexit hook;
per-test environments clean up through Drop. A full 50-test run leaves zero
stray processes.

RPC timeouts. Several tests mine 101 blocks in a single generatetoaddress,
and a dozen of them run at once, so requests sit behind a lot of work. The test
RPC clients are built by hand with a 300s transport timeout instead of jsonrpc's
15s default, and the node runs with -rpcthreads=32 -rpcworkqueue=1024.

NB_CONFIRMATIONS=6 is load-bearing. The cooperative_close tests fail at
the default of 3. This is pre-existing — verified against the current master
code pointed at an external node — and CI already sets it. It is now documented
in the README so local runs match.

CI cache. The matrix jobs now restore target/debug/build alongside
target/debug/deps, because the bitcoind and electrs executables live in the
build scripts' OUT_DIR and the test binaries resolve them through a path baked
in at compile time.

Known constraints

• The macOS electrs binary published by electrsd is x86_64, so Apple Silicon
  needs Rosetta. Linux CI is unaffected.
• postgresql_embedded is pinned to 0.19 because 0.20+ raise the MSRV to 1.94,
  above this workspace's toolchain. Bumping it later means bumping the toolchain.
• Postgres binaries are downloaded on first use and cached under
  ~/.theseus/postgresql; CI caches that directory.

Verification

Run with docker down and no connection environment variables set:

suite	result
cargo test --all-features	259 passed, 0 failed
manager_execution_tests --ignored	50/50
stateless_execution --ignored	29/29
splice_execution_tests --ignored	1/1
orphaned processes afterward	0

cargo clippy -- -D warnings and cargo fmt --check are clean.

Follow-up

The stateless-dlc-creation branch adds ddk/tests/stateless_utils.rs, which
still reads ESPLORA_HOST / BITCOIND_*. It needs the same treatment when the
two branches meet — ChainContext::new should take its esplora host and RPC
client from ddk_testenv::env(). Its 29 tests pass on the shared environment, so
no per-test isolation is needed there.