webnn: Use TFLite for CPU execution on Linux

This CL primarily does two things:
1. Use TFLite backend (instead of XNNPACK) on Linux for "cpu" contexts
2. Refactor the existing Virtual Test Suite on Linux to run "cpu"
   tests now that the "gpu" workaround is no longer needed to target
   the TFLite backend

Note that - assuming TFLite-on-CPU behaves the same across platforms -
this "cpu" Virtual Test Suite and the corresponding expectations will
be re-used as other backends (or at least Windows) migrate "cpu"
context execution to TFLite in follow-up CLs

Before (on Linux):
  - using a "cpu" context targets the XNNPACK backend in the renderer
  - using a "gpu" context targets the TFLite backend in
    //services (which currently runs on CPU anyways)
After (on Linux):
  - using either a "cpu" or "gpu" context targets the TFLite backend in
    //services

This opens the door for TFLite to truly target the GPU if instructed,
and gets us one step closer to removing the XNNPACK backend

Bug: 325612086
Change-Id: I421d97b7d91b604afbfb3c8aa71b55a5bc78bd6f

webnn/conformance_tests/compute-arraybufferview-with-bigger-arraybuffer.https.any.js

@@ -1,5 +1,6 @@
 // META: title=test WebNN MLContext.compute() for ArrayBufferView created from bigger ArrayBuffer
 // META: global=window,dedicatedworker
+// META: variant=?cpu
 // META: variant=?gpu
 // META: script=../resources/utils.js