Fix use-after-free when two Env objects share one environment

[Lotu527]

The bug

EnvWrap::open() iterated the environment cache by value:

for (env_path_t envPath : envs) {   // copy
    char* existingPath = envPath.path;
    if (!strcmp(existingPath, *charPath)) {
        envPath.count++;            // increments the copy, not the cache entry
        ...

so count on the cached entry stays at 1 no matter how many Env objects
share the environment. EnvWrap::close() then decrements the real entry, sees
zero on the first close, and calls mdb_env_close() while another Env is
still using that MDB_env.

Every later use of the first Env dereferences freed memory. In our
application this showed up as a long-standing crash in mdb_txn_begin, where
the MDB_txn read out of the environment was garbage.

Reproducing it

const lmdb = require('node-lmdb');

const env1 = new lmdb.Env();
env1.open({ path: dir, maxDbs: 4 });

const env2 = new lmdb.Env();
env2.open({ path: dir, maxDbs: 4 });   // same path -> shares env1's MDB_env

env2.close();                          // frees the shared MDB_env

env1.beginTxn();                       // reads freed memory

Before the fix this dies with SIGSEGV (exit 139) on macOS and Linux; on
Windows the same code faults as EXCEPTION_ACCESS_VIOLATION_READ.

The fix

Iterate by reference so the increment lands on the cached entry. The
environment is then closed only when its last user calls close().

Test

Added test/sharedenv.js plus a Shared environments case in
test/index.test.js. It follows the existing Cluster / Threads pattern and
runs the scenario in a child process, because the failure mode is a fatal
signal rather than a catchable exception — running it inline would abort the
mocha process and silently skip every later suite.

The close handler checks signal before code, so a regression reports
child was killed by SIGSEGV instead of a confusing null exit code.

Verified both directions locally on macOS (arm64, Node 22):

• with the src/env.cpp change reverted, the new test fails with
  child was killed by SIGSEGV: expected 'SIGSEGV' to equal null
• with the fix applied, the full suite is green: 45 passing, 1 pending
  (44 passing before this PR, plus the new case)

I could not get a CI run to show up here: .github/workflows/test.yml triggers
on push only, with no pull_request trigger, so a fork PR produces no checks.
Happy to add whatever else is useful.