Move power diag tests to validate/stress; fix GPU burn power saturation

- bee-gpu-stress.c: remove per-wave cuCtxSynchronize barrier in both
  cuBLASLt and PTX hot loops; sync at most once/sec so the GPU queue
  stays continuously full — eliminates the CPU↔GPU ping-pong that
  prevented reaching full TDP
- sat_fan_stress.go: default SizeMB 0 (auto = 95% VRAM) instead of
  hardcoded 64 MB; tiny matrices caused <0.1 ms kernels where CPU
  re-queue overhead dominated
- pages.go: move nvidia-targeted-power and nvidia-pulse from Burn →
  Validate stress section alongside nvidia-targeted-stress; these are
  DCGM pass/fail diagnostics, not sustained burn loads; remove the
  Power Delivery / Power Budget card from Burn entirely

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

audit/internal/platform/sat_fan_stress.go

@@ -20,7 +20,7 @@ type FanStressOptions struct {
 	Phase1DurSec int   // first load phase duration in seconds (default 300)
 	PauseSec     int   // pause between the two load phases (default 60)
 	Phase2DurSec int   // second load phase duration in seconds (default 300)
-	SizeMB       int   // GPU memory to allocate per GPU during stress (default 64)
+	SizeMB       int   // GPU memory to allocate per GPU during stress (0 = auto: 95% of VRAM)
 	GPUIndices   []int // which GPU indices to stress (empty = all detected)
 }
 
@@ -243,9 +243,8 @@ func applyFanStressDefaults(opts *FanStressOptions) {
 	if opts.Phase2DurSec <= 0 {
 		opts.Phase2DurSec = 300
 	}
-	if opts.SizeMB <= 0 {
-		opts.SizeMB = 64
-	}
+	// SizeMB == 0 means "auto" (worker picks 95% of GPU VRAM for maximum power draw).
+	// Leave at 0 to avoid passing a too-small size that starves the tensor-core path.
 }
 
 // sampleFanStressRow collects all metrics for one telemetry sample.