Avoid RLE for software render alpha subrect copies

[meta-legend]

• I confirm that I am the author of this code and release it to the SDL project under the Zlib license. This contribution does not contain code from other sources, including code generated by a Large Language Model ("AI").

Description

Add a software-renderer copy preparation step that disables surface RLE when all of these are true:

• the texture is static
• the texture format has alpha
• the source rectangle is smaller than, or offset within, the full texture

Whole texture alpha copies can still use RLE.

Existing Issue(s)

This resolves #15958

Benchmark

I compared unmodified SDL against this branch with a small benchmark. The benchmark creates an alpha texture atlas in memory, uses the software renderer, and renders 768 transparent 16x16 subrect copies per frame for 1000 frames.

Unmodified SDL:
frames=1000
copies_per_frame=768
seconds=1.805
fps=553.91

With my fix:
frames=1000
copies_per_frame=768
seconds=0.406
fps=2464.84

[slouken]

Merged, thanks! I'm leaving this in 3.6 for now, in case there are performance regressions with other use cases.