fix: load kernel modules via host kmod on non-container OS

[BSWANG]

Problem

On nodes whose host kmod is newer than the agent container's, modprobe erdma fails with:

modprobe: ERROR: could not insert 'erdma': Exec format error

Root cause: module load (modprobe/rmmod for erdma, smc, nvidia-peermem) ran inside the agent container via containerExec. The container's kmod (e.g. Alibaba Linux 8, kmod-25, +XZ +ZLIB only) cannot decompress ZSTD-compressed modules shipped by newer hosts (e.g. Ubuntu 24.04, kmod with +ZSTD). The module files live in the host's /lib/modules (a hostPath mount) and load into the shared host kernel, so they must be decompressed by the host kmod.

Fix

Add a moduleExec() helper that selects the exec context for kernel-module operations:

• regular OS → hostExec (enters host mount/cgroup namespaces via nsenter, using the host's ZSTD-capable kmod)
• container-optimized OS (lifsea, already runs with host userspace) → containerExec

Route modprobe erdma, EnsureSMCR (modprobe smc) and nvidia-peermem loading through it in the default, compat and ofed drivers.

This also:

• aligns the load step with the install step, which already ran on the host via hostExec for non-container OS;
• collapses the duplicated host/container selection in the OFED driver into moduleExec().

No new privileges are required — hostExec (nsenter -t 1, hostPID, SYS_ADMIN, hostPath /lib/modules) was already used by the install step and the OFED driver on these same nodes.

Test

• GOOS=linux / GOOS=darwin build, go vet, and go test ./internal/drivers/ all pass.