Add grouped-query attention with parity tests against multi-head

GQA/MQA via GPTConfig.n_kv_head; the cache stores only the KV heads. Attention
also accepts per-row positions and a key padding mask, used by batched decoding.

tests/helpers.py

@@ -0,0 +1,19 @@
+import torch
+import torch.nn as nn
+
+from microgpt.model import GPT, GPTConfig
+
+
+def make_model(cfg: GPTConfig, seed: int = 0, std: float = 0.3) -> GPT:
+    """Random model with *confident* outputs.
+
+    The default init (std 0.02) gives near-uniform logits, where greedy argmax can flip on
+    float noise and tests would compare ties. A wider init makes distinct, stable argmaxes.
+    """
+    torch.manual_seed(seed)
+    model = GPT(cfg)
+    with torch.no_grad():
+        for mod in model.modules():
+            if isinstance(mod, (nn.Linear, nn.Embedding)):
+                mod.weight.normal_(0.0, std)
+    return model.eval()