#1734 Monitor: the PCIe tile asks for pcie_rx_mb, but amdgpu only exposes a total

Beschreibung

`serve/telemetry.py` fills five PCIe readings from NVML (`_Nvml.read()`: `pcie_gen`, `pcie_gen_max`, `pcie_width`, `pcie_rx_mb`, `pcie_tx_mb`) while `_Amd.read()` reports none of them, so on Linux the Monitor's PCIe tile stays empty for AMD cards.

The link half of that is unambiguous and we have a patch for it: the narrowest and slowest hop between the root port and the card (a card that negotiates Gen4 on its own hop but sits behind a Gen3 root port really runs at Gen3).

The rate half is the question. amdgpu exposes one file, `pcie_bw`, as three integers: messages received, posted requests sent, and the maximum payload size. The kernel documents those three values and calls their product an estimate. Measured on an MI50 32 GB: idle reads `50 0 256`, and during a long prompt `4571 4709165 256` - the two counters differ by three orders of magnitude, because a card pulling experts over PCIe is the side that *posts* the requests and receives the completions.

`_Nvml`'s `pcie_rx_mb` is receive-only bytes, and the tile's own subtitle says "to GPU X MB/s". So filling that name from `pcie_bw` would put a combined figure (or a request count times payload) under a label that claims one direction.

What would you prefer: fill `pcie_rx_mb` from the received counter alone, add a separate combined field, or change the tile's wording so the number is not directional? We can do any of them - we just did not want to guess.