feat(device): optimize parse_dtb with a generic hash map

[daniel-rossier]

Closes #85

Problem

parse_dtb() matched every device tree node against the registered drivers with a nested loop: for each of the N nodes it walked the whole list of M drivers looking for a matching compatible string, once per initcall level — an O(N·M) scan.

Change

• Introduce a generic string-keyed hash map (lib/hashmap.c, include/hashmap.h): open addressing with linear probing, automatic growth, copied/owned keys, void * values. A reusable container other subsystems can rely on.
• Use it in parse_dtb() to index the drivers by their compatible string. The map is built once in O(M); the device tree is then scanned a single time and each node is matched in O(1), bringing the parsing down to O(N + M).
• Matched nodes are queued per initcall level during the single scan and only initialized afterwards — core drivers before postcore ones and in device tree order within each level — so the initialization order is unchanged.
• A single scratch dev_t is reused while probing nodes instead of being allocated and freed for every visited node.

Behaviour

Initialization order is identical. The only theoretical difference is a compatible registered at both CORE and POSTCORE: the old code initialized such a node twice (a latent double-init), the new code once at core level. This cannot happen with the current drivers — the CORE and POSTCORE compatible sets are disjoint.

Verification

• Builds cleanly (virt64_defconfig); device.o and hashmap.o compile with no new warnings; clang-format reports no diff on the touched files.
• Host equivalence test comparing the old O(N·M) and new O(N+M) matching: identical init order on all realistic cases (interleaved DT, postcore-only, disabled/absent nodes, no drivers).
• Host functional test of lib/hashmap.c: 500 inserts triggering several rehashes, exact-pointer lookups, overwrite semantics, copied-key ownership.