FBig add/sub rounding fixes, BMI2 addmul kernels, rand 0.9/0.10 support

[cmpute]

Summary

A batch of correctness fixes and performance work across dashu-int and dashu-float, plus optional support for rand 0.9 / 0.10. Several of the float fixes were surfaced and are now permanently
guarded by a differential fuzz test.

dashu-float add/sub correctness

Four rounding defects in the limited-precision alignment/rounding path:

• Severe cancellation — when the smaller operand reached into the larger operand's significant digits, the alignment path could collapse a genuine difference (e.g. 1.00 − 0.99999999 at p=3
  returned 0 instead of 1e-8). Such cases now form the exact difference at full width and round once.
• Spurious ULP on a negligible operand — the sticky was positioned at precision − digits, letting it land on a rounding tie (e.g. 1 + 2⁻¹⁰⁰ at p=10 returned 513·2⁻⁹ instead of 1). It is
  now placed at the operand's real magnitude.
• Window-edge boundary — the cancellation guard fired only on strict overlap, missing the edge case (0.5 − 0.4375 at p=1 returned 0). It now fires on >=.
• Context::sub(0, b) under directed modes — the zero-left path did -round(b) instead of round(-b), mis-rounding under the asymmetric modes Up/Down. It now rounds the negated operand
  directly.

By design, an inexact addition/subtraction may now carry a single guard digit (up to precision + 1 digits); this is documented in AGENTS.md.

dashu-int performance

• Basecase addmul/submul-2 kernels (add_mul_dword_same_len_in_place, sub_mul_dword_same_len_in_place) dispatch at runtime to a BMI2 build on x86-64 + std (flag-free mulx widening
  multiply, unrolled loop), ~4–5% faster in isolation. The portable path, no_std, and other targets are unchanged.
• Non-power-of-2 radix formatting preallocates radix_powers/big_chunks instead of growing them push-by-push.
• New Performance guide page (guide/src/performance.md) covering target-cpu=native and related build flags.

rand 0.9 / 0.10 support

New optional rand_v09 (rand 0.9) and rand_v010 (rand 0.10) features across dashu-int/-float/-ratio (plus root forwarding), mirroring rand_v08. The v08 source files are renamed to
rand_v08.rs for parallel naming, with a rand re-export alias preserving the existing dashu::*::rand paths. The default rand feature stays rand_v08.

rand 0.10 requires Rust 1.85; it's optional and is not enabled by the MSRV (1.68) CI path (which uses the explicit rand→rand_v08 feature set), so the MSRV is unchanged. rand 0.9 (MSRV
1.63) is MSRV-compatible.

Tests

• Consolidated the standalone add_sub_oracle example into the fuzz differential test: significands drawn directly via UniformBits, a deterministic small-precision sweep, and a
  guard-digit-tolerant oracle (accept rounding to p or p+1).
• Added a regression test for the Context::sub(0, b) directed-rounding fix.

Cleanup

Removed an unused BMI2 addmul-2 micro-benchmark and a stale $impl-sharing TODO.