#2559 docs: STRATA_GR_V3=1 runs only where a block has 80 KB of shared memory; on Turing, Pascal and AMD it does nothing (#2524)

Description

Closes #2524 (the reporter asked for the switch's description to say which cards it applies to).

**What it changes.** One paragraph in `docs/DETAILS.md`, after the list of opt-in switches: what `STRATA_GR_V3=1` does (the hyper-connection read in two kernels, another summation order, so answers can differ in the last bits), what it needs and where it does nothing:
- it needs 80 KB of shared memory per block: `kFusedGrMaxT` (8) x `N` (2,560) floats = 81,920 B, the check in `src/kernels/cuda/fused_gr.cu` (`need1 <= cudaDevAttrMaxSharedMemoryPerBlockOptin`);
- NVIDIA from Volta on has that (96 KB and up), except Turing (RTX 20, GTX 16: 64 KB);
- on Turing, Pascal and every AMD card (64 KB, the same code through `hip_compat`), only the two-half split would fit, which failed its parity check on Turing (#375) and is not used: the engine logs `strata: STRATA_GR_V3=1 needs the two-half split on this card, which fails its checks (#375): the default read is used` and keeps the default read; #2524 measured no change on an RX 7900 XTX;
- the one measurement in the docs: 4-6% slower on an Intel Arc B70 (SYCL port, `docs/INTEL.md`).

**What it leaves alone.** The switch, its code and every other line. No new measurement: the numbers are the code's sizes, #2524's report and INTEL.md's.

**How it was checked.** Each statement against `fused_gr.cu` (the size check and the fallback message), `include/strata/kernels/fused_gr.hpp` (`kFusedGrMaxT = 8`), the HIP attribute mapping and `git tag --contains` for 0.1.31 (the first release with the switch, already opt-in). A sub-agent reviewer checked every claim; its additions (Pascal and GTX 16 fall back too; #375 was seen on Turing) are in.

🤖 Generated with [Claude Code](https://claude.com/claude-code)