Implement 2x MSAA hack for Gen6/7

[chrisforbes]

We can do a 2x MSAA mode via some hackery:

• Program hardware in 4x mode.
• Set sample positions to -0.25 -0.25 | 0.25 0.25 | -0.25 -0.25 | 0.25 0.25 (doubled up)
• If IMS or UMS, then restrict sample mask to 0x3. [If CMS, then DONT do this, it's harmful].
• Adjust blorp resolves. Gen6 uses sample message since it does a resolve correctly for 4x. This will need to be changed to 2 ld [assuming we are masking]. Gen7 does custom resolve shaders with ld_mcs+ld2dms anyway, so no real change other than having to make sure we use the logical sample count, not the physical count.

IMS perf impact depends a lot on cache behavior. If we avoid pulling in half the cache lines, we probably win big. Hardware prefetching might break us though.

UMS perf impact should be a clear win (layers 2/3 are never touched). CMS impact should be more of a win as triangle sizes get smaller.

[chrisforbes]

Note on CMS harm: we want full coverage to translate to a single live sample value in slot 0, plus MCS bits for all samples are in slot 0.

If we masked off two samples, then the hw has to keep whatever junk was there, so MCS bits are split even for fully covered pixels. This means we can't do the all-samples-in-slot0 optimization in the resolve.

[chrisforbes]

There are a few other wrinkles:

• Alpha-to-coverage will overfill, unless it takes the programmed sample mask into account. Some w/a are possible for this.
• ARB_sample_shading will interact badly if not setting the sample mask [we'll shade the same samples twice]. Possible w/a is to discard the useless dispatches in the fs, or to mask even for CMS if sample shading is in use.