docs: clarify RISC-V RVV target selection and GCC 14+ requirement for ZVL128B/ZVL256B

[trg-rgb]

Summary

Add two notes to the RISC-V section of the build instructions, before
the per-target entries, addressing a documented source of user confusion:

1. Target selection: RISCV64_GENERIC is intentionally scalar by
   design — Makefile.riscv64 appends a scalar march override that
   takes precedence over any -march=rv64gcv flag. Users who supply
   the V extension without changing TARGET get a silently scalar build
   with no diagnostic. The correct targets for RVV 1.0 hardware are
   RISCV64_ZVL128B / RISCV64_ZVL256B, as documented in the design
   rationale at Updated RISC-V vector support for review.  #3808.
2. Compiler floor: GCC 14+ is required for the current _rvv.c
   kernel set. GCC 13 builds complete and produce a library but routines
   using segmented load/store intrinsics (__riscv_vsseg*) fall back to
   scalar silently. Functional tests pass; only disassembly detects this.
   Verified on OpenBLAS 0.3.33: GCC 13 (scalar fallback), GCC 14
   (~12,691 RVV opcodes), GCC 15 (~14,355 RVV opcodes).
   Both points are consistent with existing upstream documentation; this PR
   surfaces them at the TARGET selection step where a user would first
   encounter the choice.

Related: #3808
Community investigations: clusterchallenge/Hardware-Abstraction-Layer-Transitional-Libraries#25 (trg-rgb), clusterchallenge/Hardware-Abstraction-Layer-Transitional-Libraries#23 (Vaibhav805)

Co-authored-by: Vaibhav805 Vaibhav805@users.noreply.github.com

[martin-frbg]

Thanks - actually the GENERIC or ARCH_GENERIC target maps to the same non-vectorized plain C code on all platforms - it will result in a working build of libopenblas on cpus that have no specific optimized support, but is likely to be only a little faster than the original "netlib" Reference BLAS.

Maybe the compiler requirements for RISCV_ZVL... would be better placed in the "Building from source" documentation, probably near the end where there are architecture-specific notes for other platforms already ? The documentation is still kind of patchy, but I fear README.md is getting too bloated for something that is displayed on the github page by default.
(My experience with the RISCV platform suggests one would normally want to use the latest stable compiler version to keep up with the ever-changing extensions - hopefully we're moving past this phase now)

[trg-rgb]

Thanks @martin-frbg — that GENERIC framing is right, and I've reworded.

v2 (just pushed):

• Moved the target-selection details and compiler-requirements note out of README.md and into a new ### RISC-V subsection in docs/install.md, placed inside ## Building from source between the MIPS and FreeBSD sections, so it lives with the other per-platform build notes.
• README.md now keeps a single one-line pointer to the new section at the spot where users encounter the RISC-V build commands, so the signposting is still there without bloating the page that GitHub renders by default.
• Reworded the GENERIC framing per your correction — it's now positioned as the general "GENERIC / ARCH_GENERIC maps to plain-C on all architectures" property, with the Makefile.riscv64 -march override called out as the RISC-V-specific compounding factor.

The earlier CI failure (RISCV64_GENERIC runner, exit 140) looks unrelated to this docs-only change; happy to look more closely if it persists on this push.