perf: keep requests off the main thread for type-filtered and observer-only webRequest listeners - #53240

…r-only webRequest listeners

The IO-thread gate from #53115 sent every renderer request through the
UI-thread ProxyingURLLoaderFactory as soon as the session had any
webRequest listener, and the proxy itself wrapped every request in an
InProgressRequest for the same reason: a listener may modify the
request and re-runs on every redirect leg, so its URL filter cannot be
applied up front. Two things can be decided up front though:

- A request's resource type never changes across redirects, so the
  listeners' existing `types` filter is honored before the request
  reaches the UI thread: a blocking listener (onBeforeRequest,
  onBeforeSendHeaders, onHeadersReceived) registered with
  types: ['mainFrame', 'subFrame'] no longer pulls images, scripts and
  fetches through the main process.
- onSendHeaders, onBeforeRedirect, onResponseStarted, onCompleted and
  onErrorOccurred cannot change anything. When only such listeners
  match a request's type, it goes straight to the network service and a
  pass-through URLLoaderClient on the IO thread reports the redirect,
  response and completion to api::WebRequest afterwards; the listeners
  get the same details objects (including onSendHeaders per redirect
  leg), the renderer never waits for them.

InterceptState holds two resource-type masks instead of a boolean,
recomputed whenever a listener is added or removed, and both the gate
and the proxy's pass-through path consult them. Linux x64, page with
150 subresources from a local server while the main process is busy for
100 ms after commit: onCompleted + onErrorOccurred listeners 160 -> 55
ms (no listeners: 54 ms); fetch() round trip 0.59 -> 0.46 ms; main
process CPU per request 335 -> 244 us. With onBeforeRequest /
onHeadersReceived limited to frame types, subresources take the direct
path as well (fetch 0.61 -> 0.42 ms, 321 -> 133 us per request); the
document itself is still intercepted, so that page's load event lands
at 108 ms under the same load.

Co-authored-by: Shelley Vohr <shelley.vohr@gmail.com>