feat(grpc): native multi-backend failover via grpc-js (#144)

Support a comma-separated list in SW_AGENT_COLLECTOR_BACKEND_SERVICES so the
agent fails over between OAP backends, implemented with grpc-js's own resolver
and load balancing rather than an application-level channel state machine.

Transport
- A single address keeps the built-in `dns:` resolver, so one hostname with
  multiple A/AAAA records is load-balanced and periodically re-resolved.
- Multiple addresses register a small `sw-static` resolver (the built-in
  `ipv4:`/`ipv6:` schemes accept IP literals only) and use `pick_first` with
  `shuffleAddressList`, which spreads agents across backends while keeping the
  channel authority stable.
- Reconnection uses grpc-js exponential backoff (1s to 30s); the agent no longer
  polls channel state or re-selects backends itself.
- No client keepalive: the agent's own reporting traffic is the liveness signal,
  and a stock OAP (grpc-java defaults) answers unsolicited pings with
  GOAWAY ENHANCE_YOUR_CALM.
- Retry is scoped to ManagementService/reportInstanceProperties, the only
  idempotent unary call. Trace and meter uploads are never retried at the gRPC
  layer, so a replayed stream cannot duplicate segments.

Reporting semantics
- Best effort: a failed report discards its batch and is never re-sent, so
  statistics cannot be double-counted. Discards are logged with a running total.
- flush() is a bounded single attempt, not a delivery guarantee. It forces out
  pending data first (which has not been attempted yet), then waits on any prior
  in-flight report with the remaining budget, so a stuck RPC cannot stall a host
  such as AWS Lambda.
- Authentication failures keep the channel and emit a throttled error pointing at
  SW_AGENT_AUTHENTICATION; switching backends cannot fix a cluster-wide token.
- A backend that is reachable but unresponsive is not failed over: failover
  targets unreachable backends, not functional health.

Behavior changes
- Host http_proxy/https_proxy no longer affect the OAP uplink
  (grpc.enable_http_proxy=0); agent to OAP via HTTP proxy is not supported.
- Default SW_AGENT_LOGGING_LEVEL is now `warn`, so connection and data-loss
  diagnostics are visible without extra configuration.
- Under TLS, all backends must present certificates sharing SANs because
  verification follows the channel authority; one DNS name with multiple records
  is preferred for HA.

Docs and tests
- docs/en/setup/configuration.md and docs/en/advanced/troubleshooting.md cover the
  address forms, the diagnostics and the best-effort contract.
- Unit tests for the resolver, channel manager, report coalescing and both senders,
  plus a remote-e2e case that stops the active mock collector and asserts reporting
  continues on the standby.

No new configuration options are introduced.