docs(core): add docstrings and unit tests for SwiGLU block

- docs: rewrite and expand docstrings for SwiGLU class and forward method (motivation, math, architecture, usage, references to LLaMA/Mistral/PaLM)
- test: add unit tests for SwiGLU (shape, dtype, gradients, output range, fp16 support, reproducibility)
- strictly doc/tests, no logic or API changes

This improves transparency and reliability for gated FFN blocks in transformer architectures.

llm/tests/core/test_swi_glu.py

@@ -0,0 +1,39 @@
+import torch
+import pytest
+from llm.core.swi_glu import SwiGLU
+
+def test_swiglu_shape_and_dtype():
+    swiglu = SwiGLU(emb_size=32, dropout=0.1)
+    x = torch.randn(4, 10, 32)
+    y = swiglu(x)
+    assert y.shape == x.shape
+    assert y.dtype == x.dtype
+
+def test_swiglu_forward_range():
+    swiglu = SwiGLU(emb_size=16, dropout=0.0)
+    x = torch.randn(3, 7, 16)
+    y = swiglu(x)
+    assert y.abs().max() < 20
+
+def test_swiglu_gradients():
+    swiglu = SwiGLU(emb_size=8, dropout=0.0)
+    x = torch.randn(2, 5, 8, requires_grad=True)
+    out = swiglu(x)
+    loss = out.pow(2).sum()
+    loss.backward()
+    assert x.grad is not None
+    assert x.grad.shape == x.shape
+
+def test_swiglu_fp16():
+    swiglu = SwiGLU(emb_size=16, dropout=0.0).half()
+    x = torch.randn(1, 8, 16).half()
+    y = swiglu(x)
+    assert y.shape == x.shape
+    assert y.dtype == torch.float16
+
+def test_swiglu_reproducibility():
+    swiglu = SwiGLU(emb_size=8, dropout=0.0)
+    x = torch.ones(2, 4, 8)
+    y1 = swiglu(x)
+    y2 = swiglu(x)
+    assert torch.allclose(y1, y2)