Cut WebGPU first-flush compile time and scratch memory

[JimBobSquarePants]

Prerequisites

• I have written a descriptive pull-request title
• I have verified that there are no overlapping pull-requests open
• I have verified that I am following matches the existing coding patterns and practice as demonstrated in the repository. These follow strict Stylecop rules 👮.
• I have provided test coverage for my change (where applicable)

Description

Cuts the WebGPU first-flush time and the scratch memory the staged pipeline reserves. Measured on the DrawingBackendBenchmark sample, WebGPU, 100k lines, preview 1576x1075, AMD Radeon 8060S.

	before	after
first run of the sample	2716 to 3540 ms	360 to 380 ms with the driver shader cache warm, about 1150 ms on a machine's first launch
fine shader compile, cold	1670 to 1850 ms	1130 to 1160 ms
fine shader compile, warm	never cached	340 to 360 ms
scheduling shaders	about 1 s, one by one, inside the first flush	about 0.5 s, in parallel, at app start
scratch reserved	1.83 GB	1.12 GB
PTCL words the GPU writes	95.0 M	64.6 M
runs 2 and 3 of a cold process, before JIT tiering	58 to 90 ms	75 to 80 ms
steady state	20 to 40 ms	20 to 40 ms

Compile

• The scheduling pipelines compile when the device is probed, on all cores, instead of one by one inside the first flush.
• The fine pipeline compiles once, for the target's own format, starting when the target is created. The fixed list of four format variants is gone.
• compose_source evaluates the blend overlap once per call instead of inside five switch cases. Same operands per mode, bit-identical output. The shader now fits the driver's shader cache.
• The fine shader composes each paint command at one site after the command switch instead of inside every paint case. This cuts the cold compile from 1.8 s to 1.1 s.

Memory

• Stroke scratch is seeded from the outline geometry (offset edges, caps, joins) instead of the bounding-box diagonal. Crossings reserved for 11.5 M used: 24.6 M before, 13.65 M after.
• CMD_FILL goes from 9 words to 6, CMD_SOLID from 5 to 2, CMD_COLOR from 4 to 3. The interest rectangle and the colour are referenced by offset into the info and scene streams.
• The PTCL seed is a per-draw bound: one fill command plus that draw's paint command per tile crossing.

Verification

• No new tests. The existing 404 WebGPU tests, which include the blend-mode, gradient, clip, and 100k-line goldens, pass on net10.0 and net11.0.
• No overflow, no retry, no chunking on the 100k scene; the GPU allocator counters were read back to confirm.

[codecov]

Codecov Report

❌ Patch coverage is 95.47739% with 9 lines in your changes missing coverage. Please review.
✅ Project coverage is 79%. Comparing base (d1daf50) to head (620eae3).

Files with missing lines	Patch %	Lines
...c/ImageSharp.Drawing.WebGPU/WebGPUSceneDispatch.cs	93%	3 Missing and 2 partials ⚠️
...rc/ImageSharp.Drawing.WebGPU/WebGPUSceneEncoder.cs	97%	1 Missing and 2 partials ⚠️
src/ImageSharp.Drawing.WebGPU/WebGPURuntime.cs	66%	0 Missing and 1 partial ⚠️

Additional details and impacted files

@@          Coverage Diff           @@
##            main    #421    +/-   ##
======================================
  Coverage     79%     79%            
======================================
  Files        232     232            
  Lines      27392   27516   +124     
  Branches    3119    3128     +9     
======================================
+ Hits       21814   21934   +120     
- Misses      4540    4542     +2     
- Partials    1038    1040     +2     

Flag	Coverage Δ
unittests	79% <95%> (+<1%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.