fix(apu-memory-tuner): detect GPU when rocminfo lists Name before Device Type

[urbantech]

Problem

apu-memory-tuner's detect_platform.py misclassifies a supported AMD APU as "No AMD APU detected" (exit 2) on the exact hardware the skill targets.

Root cause is in _gfx_target_from_rocminfo(). It only captures the agent's Name: and Marketing Name: after it has already seen a Device Type: GPU line (in_gpu_agent gate). But rocminfo prints those two lines before Device Type: inside each agent block, so the gate is never satisfied for the lines that matter and the function returns ("", ""). Detection then falls through to is_apu = False.

Reproduction (Ryzen AI MAX+ 395 / Strix Halo, gfx1151 — RDNA3.5)

rocminfo field order within the GPU agent block:

Agent 2
  Name:                    gfx1151
  Marketing Name:          Radeon 8060S Graphics
  Vendor Name:             AMD
  Device Type:             GPU        <-- comes AFTER Name / Marketing Name

Before this change:

GFX target:        unknown
Generation:        unknown
APU:               False
Supported by skill: NO
  - No AMD APU detected. ...
# exit 2

Fix

Buffer the current agent block's candidate gfx target + marketing name, and commit them the moment a Device Type: GPU line confirms the block is a GPU. Reset the buffer on each Agent N header. The startswith("gfx") guard still skips the ISA sub-block Name: lines (amdgcn-amd-amdhsa--gfx1151).

After:

GPU:               Radeon 8060S Graphics
GFX target:        gfx1151
Generation:        rdna35
APU:               True
Supported by skill: YES
# exit 0

No behavior change on hardware that already detected correctly (the sysfs fallback and all other fields are untouched); this only fixes the rocminfo path that was silently returning empty.