feat(PPO_Bot): network forward pass + action mapping (#14)

Two-hidden-layer MLP actor-critic (42→64→64→5/1) with stochastic
actorForward, logStd floor at -3, and BotAction mapper wired into
the run() loop. Assert-based test suite covers shapes, finiteness,
logStd collapse, and all action range bounds.

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

PPO_Bot/actions.nim

@@ -0,0 +1,33 @@
+## actions.nim — map raw network output to Tank Royale bot commands.
+
+import arraymancer
+import std/math
+
+type
+  BotActions* = object
+    targetSpeed*:  float32
+    turnRate*:     float32
+    gunTurnRate*:  float32
+    shouldFire*:   bool
+    firePower*:    float32
+
+proc mapActions*(rawActions: Tensor[float32], currentSpeed: float32, gunHeat: float32): BotActions =
+  ## rawActions: [5] tensor from actorForward.
+  ## Ranges:
+  ##   targetSpeed  ∈ [-8, 8]
+  ##   turnRate     ∈ [-(10 - 0.75|v|), (10 - 0.75|v|)]
+  ##   gunTurnRate  ∈ [-20, 20]
+  ##   firePower    ∈ [0.1, 3.0]
+  let r0 = tanh(rawActions[0].float64).float32
+  let r1 = tanh(rawActions[1].float64).float32
+  let r2 = tanh(rawActions[2].float64).float32
+  let r3 = tanh(rawActions[3].float64).float32
+  let r4 = rawActions[4]
+
+  result.targetSpeed  = r0 * 8.0'f32
+  result.turnRate     = r1 * (10.0'f32 - 0.75'f32 * abs(currentSpeed))
+  result.gunTurnRate  = r2 * 20.0'f32
+  let fireDecision    = r3
+  result.shouldFire   = fireDecision >= 0.0'f32 and gunHeat <= 0.0'f32
+  # sigmoid(r4) * 2.9 + 0.1 → [0.1, 3.0]
+  result.firePower    = (1.0'f32 / (1.0'f32 + exp(-r4.float64).float32)) * 2.9'f32 + 0.1'f32