fix(linear): add bias once, not once per SIMD lane

acc was seeded with @splat(bias) and reduced with .Add, so every
linear output carried 7 extra copies of its bias. conv2d was unaffected
(bias via memset), which is why the trunk matched zig-solver to 5e-6
while fc/head activations diverged — the 81% labels regression.

Also adds a scalar tail for in_n < 8 (pose_fc has in_n=4) and a
known-answer selftest (test_linear).

👾 Generated with [Letta Code](https://letta.com)

Co-Authored-By: Letta Code <noreply@letta.com>

examples/test_linear.zig

@@ -0,0 +1,29 @@
+//! Minimal known-answer test for layers.linear.
+const std = @import("std");
+const crucible = @import("crucible");
+
+pub fn main(init: std.process.Init) !void {
+    const alloc = init.gpa;
+
+    // in_n = 8, out_n = 1, bias 0. w = 1..8, x = 1..8 -> expect 204.
+    var w: [8]f32 = .{ 1, 2, 3, 4, 5, 6, 7, 8 };
+    var x: [8]f32 = .{ 1, 2, 3, 4, 5, 6, 7, 8 };
+    var b: [1]f32 = .{0};
+    var out: [1]f32 = undefined;
+    crucible.layers.linear(&x, &w, &b, &out);
+    std.debug.print("case1 (expect 204): {e}\n", .{out[0]});
+
+    // bias 100
+    b[0] = 100;
+    crucible.layers.linear(&x, &w, &b, &out);
+    std.debug.print("case2 (expect 304): {e}\n", .{out[0]});
+
+    // in_n = 4 (the pose_fc case!) w=1..4 x=1..4 -> 30
+    var w4: [4]f32 = .{ 1, 2, 3, 4 };
+    var x4: [4]f32 = .{ 1, 2, 3, 4 };
+    var b1: [1]f32 = .{0};
+    var out1: [1]f32 = undefined;
+    crucible.layers.linear(&x4, &w4, &b1, &out1);
+    std.debug.print("case3 (expect 30): {e}\n", .{out1[0]});
+    _ = alloc;
+}