mm: only hugepage-align mappings of hugepage-aligned length

This is consistent with Linux after d4148aeab4124 "mm, mmap: limit THP
alignment of anonymous mappings to PMD-aligned sizes".

In #12804, when PyTorch / libcuda / the Nvidia userspace driver allocates CPU
memory, it relies specifically on Linux's mmap address selection behavior (and
the assumption that no other threads are doing mmap() / munmap()) in order to
get (in this case) a mapping aligned to one page before a hugepage boundary.
After this CL, this looks like:

```
I0401 23:24:16.756773       1 strace.go:576] [   1:   1] python3 E mmap(0x0, 0x11fff000, 0x0, MAP_PRIVATE|MAP_ANONYMOUS, 0xffffffffffffffff (bad FD), 0x0)
I0401 23:24:16.756782       1 strace.go:614] [   1:   1] python3 X mmap(0x0, 0x11fff000, 0x0, MAP_PRIVATE|MAP_ANONYMOUS, 0xffffffffffffffff (bad FD), 0x0) = 139925074735104 (0x7f42d85ff000) (3.624µs)
I0401 23:24:16.756790       1 strace.go:564] [   1:   1] python3 E munmap(0x7f42d85ff000, 0x1a01000)
I0401 23:24:16.756796       1 strace.go:602] [   1:   1] python3 X munmap(0x7f42d85ff000, 0x1a01000) = 0 (0x0) (1.436µs)
I0401 23:24:16.756801       1 strace.go:564] [   1:   1] python3 E munmap(0x7f42ea000000, 0x5fe000)
I0401 23:24:16.756805       1 strace.go:602] [   1:   1] python3 X munmap(0x7f42ea000000, 0x5fe000) = 0 (0x0) (598ns)
<driver ioctls>
I0401 23:24:16.758206       1 strace.go:576] [   1:   1] python3 E mmap(0x0, 0x10001000, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_ANONYMOUS, 0xffffffffffffffff (bad FD), 0x0)
I0401 23:24:16.758213       1 strace.go:614] [   1:   1] python3 X mmap(0x0, 0x10001000, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_ANONYMOUS, 0xffffffffffffffff (bad FD), 0x0) = 139924833562624 (0x7f42c9fff000) (1.983µs)
...
cpu.data_ptr() = 0x7f42c9fff040
7f42c9fff000-7f42da000000 rw-p 00000000 00:00 0
7f42da000000-7f42ea000000 ---p 00000000 00:00 0
```

This is consistent across executions in runc and (after this CL) gVisor. Before
this CL, this doesn't work in gVisor:

```
I0401 23:06:38.431128       1 strace.go:576] [   1:   1] python3 E mmap(0x0, 0x10001000, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_ANONYMOUS, 0xffffffffffffffff (bad FD), 0x0)
I0401 23:06:38.431136       1 strace.go:614] [   1:   1] python3 X mmap(0x0, 0x10001000, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_ANONYMOUS, 0xffffffffffffffff (bad FD), 0x0) = 140266751131648 (0x7f9265e00000) (2.221µs)
...
cpu.data_ptr() = 0x7f9265e00040
7f9265e00000-7f9275e01000 rw-p 00000000 00:00 0
```

Fixes #12804

PiperOrigin-RevId: 893266800

Author: nixprime

pkg/sentry/mm/vma.go

@@ -195,10 +195,10 @@ func (mm *MemoryManager) findAvailableLocked(length uint64, opts findAvailableOp
 		return 0, linuxerr.ENOMEM
 	}
 
-	// Prefer hugepage alignment if a hugepage or more is requested and the vma
-	// will actually be eligible for hugepages.
+	// Prefer hugepage alignment if length is hugepage-aligned and the vma will
+	// actually be eligible for hugepages.
 	alignment := uint64(hostarch.PageSize)
-	if length >= hostarch.HugePageSize && opts.Private && !opts.GrowsDown && !opts.Stack {
+	if hostarch.IsHugePageAligned(length) && opts.Private && !opts.GrowsDown && !opts.Stack {
 		alignment = hostarch.HugePageSize
 	}