Improve Linux backend: process_rw partial-transfer fix, state(), stable module handles, envar addresses

[killerra]

Summary

This PR closes three gaps in the Linux backend relative to the Windows backend, all in src/linux/process.rs:

• Implement Process::state() (previously always ProcessState::Unknown): reads /proc/<pid>/stat. Zombie (Z) and dead (X/x) states map to ProcessState::Dead with exit_code when the kernel exposes it (requires ptrace-read permission, kernel >= 3.5); a vanished PID maps to Dead(0) since the exit code is unrecoverable after the process is reaped; permission errors remain Unknown.
• Use the module base address as the module handle instead of an index into a freshly parsed /proc/<pid>/maps list. This mirrors how the Windows backend uses HMODULE and removes the race where mappings changing between module_address_list_callback and module_by_address made indices resolve to the wrong module. primary_module_address() now resolves /proc/<pid>/exe against the maps (falling back to the first file-backed mapping) instead of assuming the 0th entry.
• Report real in-process addresses for environment variables (plugin API 2): /proc/<pid>/environ is byte-identical to the env_start..env_end memory range from /proc/<pid>/stat, so each EnvVarInfo.address is derived from env_start + byte offset. environment_block_address() now returns env_start instead of the Address::NULL stub.

Testing

• cargo build, cargo clippy, cargo fmt --check clean against memflow 0.2.4.
• Smoke-tested with a throwaway example: live process reports Alive; an unreaped child reports Dead(0); primary_module_address() resolves to the actual executable mapping and round-trips through module_by_address; module list addresses match mapping bases.
• Note: the #[cfg(memflow_plugin_api = "2")] envar path compiles out against released memflow (plugin API 1), same as the pre-existing envar code in this repo, so it is not covered by local builds or CI. The changes only use the same EnvVarInfo surface as the existing Windows API-2 code plus Stat::env_start from procfs 0.15.1.

Audit follow-ups (second push)

A soundness review of the rest of the Linux backend surfaced one serious pre-existing bug plus several metadata gaps, addressed in two additional commits:

• Fix partial-transfer retry window in process_rw (src/linux/mem.rs): after a partial process_vm_readv/writev, the retry syscall resumed at the iov window offset, but result dispatch iterated the iovecs and metadata from index 0 — re-reporting already-delivered entries with wrong metadata/buffers, dropping the entries the retry actually transferred, and flagging the wrong element as failed. Any batched read spanning an unmapped hole was affected. Comes with a regression test (batched [valid, unmapped, valid] read against our own PID) that fails on the previous code.
• Decode the waitpid-style status word in state() so exit(3) reports Dead(3) rather than Dead(768); signal deaths report the negated signal number. Unit-tested.
• Derive architecture from the compile target (x86_64/x86/aarch64) instead of hardcoding x86-64 in OsInfo.arch and ProcessInfo::{sys_arch,proc_arch}, and make the module list callback emit the same arch field module_by_address keys on.
• Stable OS kernel-module handles: FNV-1a hash of the module name instead of indices into a name-sorted snapshot (which shifted on module load/unload — the same instability class this PR fixed for process modules).
• process_info_by_pid reports the real process state via the shared helper instead of hardcoding Alive.

Deferred (noted, not in this PR): CI coverage for the plugin-API-2 cfg paths, and a pidfd-based mitigation for the inherent PID-recycling race.