goke-osdrv-gk7205v200: ship the VQE engines the microphone path loads by dlopen

[openipc-ai]

The Goke half of #2361, split out because it is the only part that adds binaries and the only part with no hardware to validate it on.

Do not merge until someone runs it on a GK7205V200. Opened as a draft so the work is not lost and is ready the moment a board is available.

What it does

Identical mechanism to the HiSilicon side. libupvqe.so is only the framework — it dlopens one shared object per DSP stage when HI_MPI_AI_EnableVqe() runs, resolving names from a lib%s_%s.so template, and prints dlopen libvqe_common.so/libsecurec.so or %s failed when they are missing. Goke names them libvqe_* where HiSilicon names them libhive_*.

Installs only the stages majestic can switch on — libvqe_hpf, libvqe_anr, libvqe_agc, libvqe_common (a NEEDED of the last two, not a stage), and libvqe_record for the 48 kHz path. AEC and EQ are not exposed; the resampler is compiled into libupvqe.so itself (RES_ReSampler_* are defined there), so libvqe_res.so is never loaded.

Gated ifeq ($(OPENIPC_MAJESTIC),ultimate), and placed in files/vqe/ rather than files/lib/ because this package installs files/lib/* wholesale and there is no other way to keep 360 KB out of a lite image.

Provenance

The five libraries come from the GK7205V200 vendor SDK drop — the same drop the twenty libraries already in files/lib/ came from. Before copying these five I verified every one of those twenty is byte-identical to a file in that drop, libupvqe.so (the loader that will dlopen these) among them.

That satisfies the checklist's success criterion as written — "traces to a vendor SDK release or to a source tree the build system compiles" — but provenance is not runtime validation, which is the open question below.

What is missing

No GK7205V200 on the bench. Matching provenance and an identical dlopen mechanism are not the same as watching the enable call succeed on the part, and I am not willing to claim otherwise.

What needs running on a real board, on an ultimate image built from this branch:

1. Start majestic with audio.enabled: true, audio.srate: "8000", audio.vqe: true.
2. Confirm the log shows talk VQE on at 8000 Hz with no dlopen ... failed or create vqe fail on the console.
3. Capture http://<cam>/audio.pcm with vqe: false and again with vqe: true, same room, and compare energy in 20–80 Hz against 300–1000 Hz. The high-pass filter should cut the former by an order of magnitude and leave the latter alone.

For reference, that test on the two HiSilicon boards in #2361 gave:

board	20–80 Hz (stopband)	300–1000 Hz (speech)
hi3516ev300 (3516E200)	0.03x	0.98x
hi3516av300 (3516C500)	0.01x	1.38x

If any engine turns out to be ABI-incompatible with the shipped Goke userspace, the fix is to take the binaries from whichever SDK drop matches libupvqe.so in files/lib/ — but that is already the drop these came from.