fix(random): remixture TRNG on rand_below retry; Linux getrandom

[SashaMIT]

Summary

Two related entropy hygiene fixes in ngu/random.c:

1. _rand_below / uniform: After the first sample, rejection retries previously remixed only my_yasmarang() and never called CHIP_TRNG_32() again. Retries now remixture TRNG each iteration.
2. Linux CHIP_TRNG_32: Was mapped to glibc random() with no srandom/getrandom. Now uses getrandom(2) (fails hard on error, same pattern as the duplicate-TRNG check in my_random_bytes).

Impact

• Affects ngu.random.uniform consumers (Coldcard firmware uses this for Key Teleport / PIN scramble / web2fa digits, etc.).
• Linux/simulator path: critical if anyone generates real keys on that build; STM32/Apple paths unchanged for (2).

Test plan

• Build makefile.unix / MicroPython unix port
• Spot-check uniform still returns < mx
• Confirm Linux no longer links behaviour to unseeded random()

Elacity CodeRED Amber review (portal #51, #53). Coordinating with Coinkite on firmware-side siblings separately.

[SashaMIT]

Follow-up commit: reseed(bytes) now absorbs into full yasmarang state (pad/n/d/dat). Integer reseed(int) kept for compatibility.

Pairs with Coldcard firmware change to pass sha256d(SE1||SE2) bytes instead of truncating to 32 bits (portal #52).

[SashaMIT]

Friendly bump — related entropy hardening across libngu#65 and Coldcard firmware#697 (TRNG remix on reject, Linux getrandom, full SE hash reseed). Happy to split or retarget if you want a different shape.

[switck]

rebase, will merge

[SashaMIT]

Rebased onto current master (e9d5e80). Note: your my_random_bytes() refactor of the _rand_below retry already covers the re-mixture half of this PR, so I've rescoped to the still-live half: CHIP_TRNG_32 on Linux mapped to unseeded glibc random(), which makes drbg_seed_from_chip deterministic on Linux/sim builds. Now maps to getrandom(2), hard-failing on error.