README: fix Usage block to match published headline + new tester defaults

- Step 3: loss._lambda=1.5 -> loss._lambda=1, the lambda used for the
  paper's Table 2/3 results (and the lambda whose joint .ckpt is now
  published in WEIGHT_FILES as ninja_desc_<base_desc>_joint).
- Evaluation: drop the stale epoch=21 step=13749 example (which pointed
  at an internal eq_lr/2021-07-16_run1 path that no longer applies).
  The default invocation now relies on lemurianet_module.py's auto-
  download of the published joint checkpoint when
  padesc_checkpoint.base_dir is null. Custom-run example uses
  padesc_checkpoint.epoch=last instead.

Author: hyojinie

README.md

@@ -98,18 +98,24 @@ python -m ninjadesc.pa_desc.tools.recon_trainer \
 
 ```bash
 python -m ninjadesc.pa_desc.tools.trainer \
-    profiler=simple trainer.max_epochs=300 trainer.gpus=4 loss._lambda=1.5
+    profiler=simple trainer.max_epochs=300 trainer.gpus=4 loss._lambda=1
 ```
 
-`loss._lambda` is the privacy parameter that trades off matching utility against image-reconstruction concealment.
+`loss._lambda` is the privacy parameter that trades off matching utility against image-reconstruction concealment. The paper's headline numbers (Table 2/3) are at `loss._lambda=1`.
 
-**Evaluation** on MegaDepth:
+**Evaluation** on MegaDepth — by default the tester auto-downloads the published joint-training checkpoint for the chosen `base_desc` (`ninja_desc_<base_desc>_joint.ckpt`) when `padesc_checkpoint.base_dir` is left null:
+
+```bash
+python -m ninjadesc.pa_desc.tools.tester base_desc=sosnet
+```
+
+To evaluate your own joint-training run instead, point at its checkpoint directory:
 
 ```bash
 python -m ninjadesc.pa_desc.tools.tester \
     base_desc=sosnet \
     padesc_checkpoint.base_dir=${NINJADESC_OUTPUT_ROOT}/pa_desc_joint/<run> \
-    padesc_checkpoint.epoch=21 padesc_checkpoint.step=13749
+    padesc_checkpoint.epoch=last
 ```
 
 ## Repository layout