Fix out-of-bounds write in drawChar() classic glcdfont path

[94xhn]

Problem

The classic built-in glcdfont branch in Arduino_GFX::drawChar() sends pixels directly to writePixelPreclipped(). Its loops checked _max_text_x and _max_text_y, but not the corresponding minimum text bounds.

When a partially visible character starts above or left of the text area, negative coordinates can therefore reach canvas implementations that index their framebuffer directly and rely on the caller to have clipped them.

Fix

Add the missing _min_text_x and _min_text_y checks to the existing column and row conditions. Visible pixels continue to use writePixelPreclipped() directly; only pixels outside the top or left text boundary are skipped.

Verification

I built a host harness against the real repository Arduino_G.cpp, Arduino_GFX.cpp, and Arduino_Canvas.cpp, using a ProbeCanvas subclass only to record coordinates before forwarding each call to the real canvas implementation.

• Before the fix, setCursor(-3, -2); print("A") made 40 preclipped writes: 13 had framebuffer offsets outside [0, 4096), and 15 more had invalid coordinates that mapped to unrelated in-buffer pixels.
• With the minimum-bound checks, the same case makes 12 visible preclipped writes and zero invalid-coordinate or out-of-buffer writes.
• setCursor(-3, 5) likewise completes with 16 visible writes and zero invalid writes.
• Fully visible and default canvas-clipped framebuffer hashes match the prior bounds-safe implementation.
• A custom text bound starting at (4, 5) also clips pixels against its minimum x/y values.

The host harness compiles and runs successfully. Physical display hardware was not used.

[moononournation]

changing writePixelPreclipped to writePixel is not a good idea, please pay more effort to make it.

[94xhn]

Updated in 23a4da0. The glyph path now keeps writePixelPreclipped() and adds the missing _min_text_x/_min_text_y checks alongside the existing maximum-bound checks. The real Canvas harness reports zero invalid writes for the negative-cursor cases, while fully visible and canvas-clipped framebuffer hashes remain unchanged. I also corrected the PR description to match the measured results.