fix(shell): improve reliability — keepalive, error classification, agent-timeout detection

[Guimove]

Context

Customer bug (Jonathan Petitcolas): `qovery shell` WebSocket connections drop with "Deadline of 30s exceeded for receiving agent response" because the K8s namespace lookup + pod list + exec can take longer than the hardcoded 30s gateway timeout.

Companion MR on rust-backend: qovery/backend/rust-backend!593

Changes

Keepalive

• Replace unbounded read deadline with ping/pong keepalive (`PingInterval=30s`, `ReadTimeout=75s`).
  Idle-but-healthy connections now survive indefinitely; only truly dead connections are detected and closed.

WebSocket error classification (`wserror.go`)

• Permanent errors (1007/1008): cancel the reconnect loop immediately — permission denied or auth rejected, retrying won't help.
• Transient errors (1011): reconnect with delay.
• Agent timeout subset (`IsAgentResponseTimeout`): 1011 errors whose text contains one of:
  • `"exceeded for receiving agent response"` — gateway DEFAULT_AGENT_RESPONSE_TIMEOUT
  • `"while connecting to pod"` — shell-agent KUBE_OPERATION_TIMEOUT (K8s exec)
  • `"while setting up port forward"` — shell-agent KUBE_PORT_FORWARD_TIMEOUT
  • `"Retry budget exhausted"` — shell-agent retry budget guard
    These show a specific "retrying…" message instead of the generic "service unavailable" one.

Port-forward

• Replace `log.Fatal` with `log.Errorf` + return so the process does not exit on a connection error.
• Improved error messages matching shell-agent timeout strings.

Tests

• `go test ./pkg` — all wserror tests pass.
• 19 tests covering `IsPermanentCloseError`, `IsInternalServerError`, `IsAgentResponseTimeout` (including wrapped errors and ordering invariant), `ServiceUnavailableMessage`.