refactor(cli): remove duplicate connection code

[kacy]

summary

when ember-client was added as a standalone crate, the connection logic
was written fresh rather than wiring the CLI to use it. the result was ~85%
duplication: both crates had their own Connection/Client struct, identical
error types, the same frame read loop, and a copy of TlsClientConfig.

this PR removes the dead copy from ember-cli and wires it to ember-client.

what changed:

• ember-client::TlsClientConfig is now public and re-exported from the crate root
• ember-cli::tls re-exports it instead of defining its own identical struct
• ember-cli gains an ember-client dependency
• all callers (repl, batch, watch, cluster, oneshot) now use Client::connect /
  Client::connect_tls, .send(), .auth(), and .disconnect()
• auth_frame in bench_conn.rs was a 4-line, single-caller helper — inlined directly
• crates/ember-cli/src/connection.rs deleted

bench_conn.rs still owns MaybeTlsStream and the raw TLS connect path for
pipelining — that code is benchmark-specific and stays.

what was tested

• cargo build -p emberkv-cli — clean, zero warnings
• cargo clippy -p emberkv-cli -- -D warnings — clean
• cargo test -p emberkv-cli — 114/114 pass
• cargo test --workspace — 78/79 integration tests pass; the failing
  cluster_moved_redirect test is a pre-existing flaky race unrelated to this change
• cargo fmt --all --check — clean

design notes

bench_conn.rs bypasses Client intentionally — it pre-serializes commands
and writes raw bytes for pipeline benchmarking. it still reads from
crate::tls::MaybeTlsStream directly, which is the right level of abstraction
for that use case.