Route VM-dictionary sf_malloc/sf_free through the kernel's kmalloc heap

Captain Bob's call after seeing the identity-heap-capacity findings
(FABRIC-3.md §X.4): the number of concurrently-running VMs is not known
in advance, and once this is a complete operating system the heap should
be able to use whatever memory is actually available -- not a hardcoded
compile-time ceiling. This was already half-built and just not wired up.

src/starkernel/vm/alloc_kernel.c previously implemented sf_malloc()/
sf_free() (platform_alloc.h's allocator abstraction -- what
vm_create_word() calls for every VM's word dictionary) as its own
isolated static 4MB arena: first-fit free list, no splitting or
coalescing. That's exactly the allocator that topped out around 6
concurrent WIREBIND-born identities, failing from fragmentation before
true capacity exhaustion (§X.4's own measurements).

Sitting right next to it, unused for this purpose: src/starkernel/memory/
kmalloc.c, the kernel's general heap. Already initialized at boot (M6,
kernel_main.c, well before any VM is ever born), reserved from real
PMM-tracked physical memory rather than a fixed array, defaults to a
2 GiB floor explicitly sized "for 256+ baby VMs" per its own comment,
overridable via the --heap= boot flag, and its free list actually
coalesces neighboring blocks on every free.

Change: alloc_kernel.c's sf_malloc()/sf_free() now delegate to
kmalloc_aligned()/kfree() instead of managing a separate arena.
sf_alloc_init() becomes a no-op (kmalloc is already initialized by the
time any VM allocation can happen, and "resetting" a heap now shared by
every kernel subsystem would be actively wrong -- confirmed no external
caller depended on its old reset semantics). sf_alloc_get_stats() reads
kmalloc_get_stats() fresh rather than shadowing byte counts locally;
alloc_count/free_count (which kmalloc.c doesn't track) stay as simple
local counters. sf_calloc()/sf_realloc() are otherwise unchanged. Kernel-
only: the hosted (non-kernel) StarForth build keeps its own separate
alloc_host.c implementation, untouched.

Verified live: replaying the exact hotplug sequence that previously
topped out at 6 identities (Zuse + 8 identities, one at a time via QMP
device_add) now succeeds for all 9, where identity 05 specifically used
to fail. Three-arch clean qemu acceptance (single Zuse device, the
standard regression case) passed on amd64, aarch64, and riscv64 -- one
aarch64 attempt hit an unrelated, already-documented one-off QEMU hiccup
(empty log, boot never progressed past firmware) and passed cleanly on
retry with no rebuild.

Not addressed here: the underlying free-list itself is still first-fit
without splitting (only coalescing changed, inherited from kmalloc.c);
per-VM dictionary sizing (shrinking what each WIREBIND VM's word set
actually needs) is a separate, still-open lever from FABRIC-3.md §X.4's
open architecture question.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014Ec88YKxxhZGG1RNnune78

This commit is contained in: