portal_devfs: don't install cdev-only poll/write proxies on anon/sock files

[lacraig2]

Problem

The issue #77 / #78 poll & write proxies (igloo_devfs_proxy_poll, igloo_devfs_proxy_write) resolve their tracking struct with container_of(inode->i_cdev, struct portal_devfs_entry, cdev) — they are char-device-only.

#78 installed them inside the shared igloo_convert_ops_to_fops(), but that converter is also called by portal_anon.c (handle_op_anonfs_create_file / sockfs) to build fops for anon inodes and sockets, which have no cdev (inode->i_cdev == NULL). As a result igloo_devfs_proxy_write hits if (!inode->i_cdev) return -ENODEV; and every write to an anon file or socket fails with ENODEV.

This surfaced as the anonfs unit tests (anonfs_test_pass/_counter/_socket) going red on every arch the moment a tree pins driver 0.0.81 (e.g. penguin #845, and it will hit anyone bumping past 0.0.80). echo 5 > /tmp/anon_counter fails, the set -eux script aborts before PASS. Pre-0.0.81 the converter set out->write = ops->write directly, so anon/sock writes worked.

Fix

Keep the shared converter on the direct Python-modeled write/poll, and install the cdev-only proxies only in the char-device create path (handle_op_devfs_create_device), where the cdev and poll_wq actually exist. The #77 blocking-poll + wake-on-write behavior for char devices is unchanged. Block devices call python_write directly via queue_rq, so they don't need the proxy either.

Net diff: one file, src/portal/portal_devfs.c.

Follow-up

Once merged and 0.0.82 is released, penguin #845 will be re-pinned 0.0.81 → 0.0.82 (restores anonfs green there).