A log line per request, and every API error logged once

A self-hosted server's logs now tell a request in one line, and an error in one more.

• One INFO line per request, request served, with its method, its path without the query, its operation id, status, size and duration and, inside a sampled trace, TraceID and SpanID. The handling request line and the error_code field are gone.
• Every error the API returns is logged once, as an api error line with the status, the operation, and the error's code, message and cause: a 5xx at ERROR, with its location, and a 4xx at INFO. A handler that panics before responding now answers 500 internal instead of closing the connection.
• Stack traces are off by default. Set instrumentation.log.errors.stack_trace: true (NEXTGEN_INSTRUMENTATION_LOG_ERRORS_STACK_TRACE=true) while debugging.
• Removing a stream from instrumentation.log.streams now drops that stream's lines, and instrumentation.log.add_source applies to every line.
• A server that fails to start logs why at ERROR, in the configured log format, so the reason also shows with instrumentation.log.level: warn.
• instrumentation.trace.trust_remote_spans: true (NEXTGEN_INSTRUMENTATION_TRACE_TRUST_REMOTE_SPANS) continues a caller's W3C traceparent trace; it was accepted before but did nothing. It trusts the caller's sampled flag over instrumentation.trace.fraction, so turn it on only behind a proxy or mesh you control.