starboard/testing: Use PBufferSurface for FakeGraphicsContextProvider.

[yell0wd0g]

FakeGraphicsContextProvider is used by NPLB tests to create a, well, Fake graphics scaffolding. However, in InitializeEGL(), it creates a real EGLSurface (i.e. connected to the actual screen), which of course is unnecessary (because we want a Fake). This CL changes the code to create an off-screen pixel buffer surface (PBufferSurface).

It also takes the chance to modernize the code: introduces constexpr, for-range loop, and std::vector (to avoid malloc-free).

I wanted to do this a while back to address nplb flakes: locally I saw eglCreateWindowSurface() failing to allocate the needed EGLSurface, and I thought it could be due to too many back-to-back EGL construct-initialize-destruct cycles. However, the flake was still there so I decided to not touch the code. But #4987, #5038 have made this change necessary.

This was tested locally on my gLinux box running:
xvfb-run ./out/linux-x64x11_devel/nplb --v=3 --gtest_filter=*SbPlayer* - 567 seconds, 2920 tests, 8 passed, 2192 skipped.

Then I run all the tests, just to be sure: xvfb-run ./out/linux-x64x11_devel/nplb --v=3

[==========] 3522 tests from 178 test suites ran. (640050 ms total)
[  PASSED  ] 591 tests.
[  SKIPPED ] 2925 tests, listed below:
.....
[  FAILED  ] 6 tests, listed below:
[  FAILED  ] SbAudioSinkTest.UpdateStatusCalled
[  FAILED  ] MultiplePlayerTest.SunnyDay
[  FAILED  ] SbMediaSetAudioWriteDurationTests/SbMediaSetAudioWriteDurationTest.WriteLimitedInput/0, where GetParam() = "beneath_the_canopy_aac_stereo.dmp"
[  FAILED  ] SbMediaSetAudioWriteDurationTests/SbMediaSetAudioWriteDurationTest.WriteContinuedLimitedInput/0, where GetParam() = "beneath_the_canopy_aac_stereo.dmp"
[  FAILED  ] VerticalVideoTests/VerticalVideoTest.WriteSamples/audio_silence_aac_stereo_dmp_video_vertical_1080p_30_fps_137_avc_dmp_output_decode_to_texture_key_system_null, where GetParam() = 40-byte object <B4-D3 4E-E3 F0-55 00-00 02-D4 4E-E3 F0-55 00-00 00-00 00-00 AB-AB AB-AB 57-16 51-E3 F0-55 00-00 57-16 51-E3 F0-55 00-00>
[  FAILED  ] VerticalVideoTests/VerticalVideoTest.WriteSamples/audio_silence_aac_stereo_dmp_video_vertical_1080p_30_fps_137_avc_dmp_output_punch_out_key_system_null, where GetParam() = 40-byte object <B4-D3 4E-E3 F0-55 00-00 02-D4 4E-E3 F0-55 00-00 01-00 00-00 AB-AB AB-AB 57-16 51-E3 F0-55 00-00 57-16 51-E3 F0-55 00-00>

In a Linux CI run, for the curious among us:

2025-03-10T16:15:30.7580398Z [----------] Global test environment tear-down
2025-03-10T16:15:30.7580522Z [==========] 3514 tests from 174 test suites ran. (171417 ms total)
2025-03-10T16:15:30.7580608Z [  PASSED  ] 589 tests.
2025-03-10T16:15:30.7580699Z [  SKIPPED ] 2925 tests, listed below:

Bug: b/384819454

[yell0wd0g]

@xiaomings and @sherryzy PTAL. Bots are still not running but changes work locally.

[sherryzy]

The code LGTM.

Looking at your summary, there are more than 2000 SbPlayer tests, only 8 of them are run. Do you know why? Also for the 6 failed tests, are they related to your change or no?

[yell0wd0g]

The code LGTM.

Looking at your summary, there are more than 2000 SbPlayer tests, only 8 of them are run. Do you know why?

They do some internal feature detection and decide to GTEST_SKIP() - some examples are protected content (not supported on Linux Cobalt yet) or hardware decoding (also not supported on Linux IIRC).

Also for the 6 failed tests, are they related to your change or no?

They are failing on ToT too -- note that I run nplb "vanilla" and that's why I see them failing, on CI they are suppressed using the cobalt/testing/filters/linux-x64x11/nplb_filter.json file.