Record check timings only when -XepPrintTimings asks for them

[vlsi]

Why

ErrorProneScanner.processMatchers opens a timing span for every matcher it runs on every AST node, and nothing reads what the spans record: ErrorProneTimings.timings() and initializationTime() have no callers in error_prone_check_api or error_prone_core, and no test asserts on either.

The spans are not cheap at that frequency. Compiling error_prone_core itself opens 14771508 of them, and Calcite's :core opens 28493608, each a HashMap lookup, two System.nanoTime calls, and a closure. Against a build whose span returns a shared no-op, alternating the two inside one JVM so each pair meets the same machine state, they cost a median of 656 ms of a 13.0 s compile and 1216 ms of a 37 s one, and on Calcite 0.42 GiB of allocation.

#6079 has the full measurements.

What

-XepPrintTimings gates the spans, so a compilation that does not pass it opens none. The same flag prints the per-check totals when the compilation finishes, so the data reaches whoever paid for it:

Error Prone ran 440 checks in 582 ms, and spent 338 ms initializing
       172 ms   29.6%  ParameterName
        27 ms    4.8%  AvoidCommonTypeNames
        24 ms    4.2%  AutoValueSubclassLeaked

Every check that ran gets a row, so the count in the header is the number of rows below it.

An embedder that reads timings() programmatically now needs the flag too. That is the part worth arguing with, and #6079 says why I think the default belongs this way round: the collection has no consumer in this repository, and the flag is what gives it one.

Making the collection itself cheap enough to leave on is a larger change, and #6079 proposes it separately.

How to verify

mvn -pl check_api,core test -Dtest=ErrorProneOptionsTest,ErrorProneJavaCompilerTest

recognizesPrintTimings and printTimingsIsOffByDefault cover the option. printTimingsReportsEveryCheckThatRan compiles a real file and asserts both a non-zero count in the header and at least one row: the header alone is printed before the rows, so a build that collected nothing would still print one. withoutPrintTimingsNoReportIsPrinted covers the default.

Each test was watched failing: deleting the option's parse case fails the first, inverting the gate in VisitorState.timingSpan fails the third on both of its assertions, and dropping the Kind.COMPILATION branch fails it on the absent header.

One limitation the tests do not cover: under Gradle the report does not reach the build log, because Gradle passes no writer to JavaCompiler.getTask and javac's NOTICE writer then goes to the compiler daemon's stderr. It arrives through any caller that supplies a writer. Emitting it as a diagnostic instead would fix that, and the same applies to what RefactoringCollection already prints; I left both alone here.